Merge pull request #10546 from chengduoZH/feature/change_pe_strategy

reyoung · web-flow · commit 9923be5d10b1 · 2018-05-14T10:30:44.000+08:00
Balance parameter_opt between cards
diff --git a/paddle/fluid/framework/details/multi_devices_graph_builder.cc b/paddle/fluid/framework/details/multi_devices_graph_builder.cc
@@ -37,20 +37,26 @@ MultiDevSSAGraphBuilder::MultiDevSSAGraphBuilder(
     const std::string &loss_var_name,
     const std::unordered_set<std::string> &params,
     const std::vector<Scope *> &local_scopes,
-    platform::NCCLContextMap *nccl_ctxs, bool use_default_grad_scale)
+    platform::NCCLContextMap *nccl_ctxs, bool use_default_grad_scale,
+    bool balance_parameter_opt_between_cards)
     : loss_var_name_(loss_var_name),
       places_(places),
       local_scopes_(local_scopes),
-      nccl_ctxs_(nccl_ctxs) {
+      nccl_ctxs_(nccl_ctxs),
+      balance_parameter_opt_between_cards_(
+          balance_parameter_opt_between_cards) {
 #else
 MultiDevSSAGraphBuilder::MultiDevSSAGraphBuilder(
     const std::vector<platform::Place> &places,
     const std::string &loss_var_name,
     const std::unordered_set<std::string> &params,
-    const std::vector<Scope *> &local_scopes, bool use_default_grad_scale)
+    const std::vector<Scope *> &local_scopes, bool use_default_grad_scale,
+    bool balance_parameter_opt_between_cards)
     : loss_var_name_(loss_var_name),
       places_(places),
-      local_scopes_(local_scopes) {
+      local_scopes_(local_scopes),
+      balance_parameter_opt_between_cards_(
+          balance_parameter_opt_between_cards) {
 #endif
   for (auto &p : params) {
     grad_names_.insert(GradVarName(p));
@@ -124,6 +130,12 @@ std::unique_ptr<SSAGraph> MultiDevSSAGraphBuilder::Build(
   // Find "send" op first for split is in front of send.
   OpDesc *send_op = GetSendOpDesc(program);
 
+  size_t cur_device_id = 0;
+  std::vector<std::unordered_set<std::string>> var_name_on_devices;
+  std::vector<std::unordered_set<std::string>> bcast_var_name_set;
+  var_name_on_devices.resize(places_.size());
+  bcast_var_name_set.resize(places_.size());
+
   bool is_forwarding = true;
   for (auto *op : program.Block(0).AllOps()) {
     if (op->Type() == "send") {
@@ -139,24 +151,47 @@ std::unique_ptr<SSAGraph> MultiDevSSAGraphBuilder::Build(
       }
       is_forwarding = false;
     } else {
-      CreateComputationalOps(&result, *op, places_.size());
+      int op_dev_id = GetOpDeviceID(var_name_on_devices, *op);
+      if (op_dev_id == -1) {  // var on all device
+        CreateComputationalOps(&result, *op, places_.size());
+      } else {
+        CreateComputationalOp(&result, *op, op_dev_id);
+        for (auto &var_name : op->OutputArgumentNames()) {
+          var_name_on_devices[op_dev_id].emplace(var_name);
+        }
+      }
       if (!is_forwarding && places_.size() > 1) {
         // Currently, we assume that once gradient is generated, it can be
         // broadcast, and each gradient is only broadcast once.
         for (auto &og : op->OutputArgumentNames()) {
           if (IsParameterGradientOnce(og, &og_has_been_broadcast)) {
-            if (IsSparseGradient(var_types, og)) {
-              CreateReduceOp(&result, og, 0);
-              CreateBroadcastOp(&result, og, 0);
+            if (balance_parameter_opt_between_cards_) {
+              CreateReduceOp(&result, og, cur_device_id);
+              var_name_on_devices[cur_device_id].emplace(og);
+              bcast_var_name_set[cur_device_id].emplace(
+                  og.substr(0, og.size() - strlen(kGradVarSuffix)));
+              cur_device_id = (cur_device_id + 1) % places_.size();
             } else {
-              InsertNCCLAllReduceOp(&result, og);
+              if (IsSparseGradient(var_types, og)) {
+                CreateReduceOp(&result, og, 0);
+                CreateBroadcastOp(&result, og, 0);
+              } else {
+                InsertNCCLAllReduceOp(&result, og);
+              }
             }
           }
         }
       }
     }
   }
 
+  // Insert BCast Ops
+  for (size_t dev_id = 0; dev_id < bcast_var_name_set.size(); ++dev_id) {
+    auto &to_bcast_set = bcast_var_name_set[dev_id];
+    for (auto &bcast_name : to_bcast_set) {
+      CreateBroadcastOp(&result, bcast_name, dev_id);
+    }
+  }
   /*
     Dependency graph has been constructed. However, there are still data
     harzaeds need to be handled.
@@ -265,6 +300,26 @@ bool MultiDevSSAGraphBuilder::IsParameterGradientOnce(
   return is_pg_once;
 }
 
+int MultiDevSSAGraphBuilder::GetOpDeviceID(
+    const std::vector<std::unordered_set<std::string>> &var_name_on_devices,
+    const OpDesc &op) const {
+  if (!balance_parameter_opt_between_cards_) {
+    return -1;
+  }
+
+  int var_dev_id = -1;
+  for (auto &var_name : op.InputArgumentNames()) {
+    if (var_dev_id != -1) break;
+    for (size_t i = 0; i < var_name_on_devices.size(); ++i) {
+      if (var_name_on_devices[i].count(var_name)) {
+        var_dev_id = static_cast<int>(i);
+        break;
+      }
+    }
+  }
+  return var_dev_id;
+}
+
 void MultiDevSSAGraphBuilder::CreateScaleLossGradOp(SSAGraph *result) const {
   for (size_t i = 0; i < places_.size(); ++i) {
 // Insert ScaleCost OpHandle
diff --git a/paddle/fluid/framework/details/multi_devices_graph_builder.h b/paddle/fluid/framework/details/multi_devices_graph_builder.h
@@ -36,13 +36,15 @@ class MultiDevSSAGraphBuilder : public SSAGraphBuilder {
                           const std::unordered_set<std::string> &params,
                           const std::vector<Scope *> &local_scopes,
                           platform::NCCLContextMap *nccl_ctxs,
-                          bool use_default_grad_scale);
+                          bool use_default_grad_scale,
+                          bool balance_parameter_opt_between_cards);
 #else
   MultiDevSSAGraphBuilder(const std::vector<platform::Place> &places,
                           const std::string &loss_var_name,
                           const std::unordered_set<std::string> &params,
                           const std::vector<Scope *> &local_scopes,
-                          bool use_default_grad_scale);
+                          bool use_default_grad_scale,
+                          bool balance_parameter_opt_between_cards);
 #endif
 
   std::unique_ptr<SSAGraph> Build(const ProgramDesc &program) const override;
@@ -60,6 +62,7 @@ class MultiDevSSAGraphBuilder : public SSAGraphBuilder {
 #ifdef PADDLE_WITH_CUDA
   platform::NCCLContextMap *nccl_ctxs_;
 #endif
+  bool balance_parameter_opt_between_cards_;
   bool use_default_grad_scale_;
 
   bool IsScaleLossOp(const OpDesc &op) const;
@@ -84,6 +87,10 @@ class MultiDevSSAGraphBuilder : public SSAGraphBuilder {
       const std::string &og,
       std::unordered_set<std::string> *og_has_been_broadcast) const;
 
+  int GetOpDeviceID(
+      const std::vector<std::unordered_set<std::string>> &var_name_on_devices,
+      const OpDesc &op) const;
+
   void InsertNCCLAllReduceOp(SSAGraph *result, const std::string &og) const;
 
   void CreateBroadcastOp(SSAGraph *result, const std::string &p_name,
diff --git a/paddle/fluid/framework/parallel_executor.cc b/paddle/fluid/framework/parallel_executor.cc
@@ -58,7 +58,7 @@ ParallelExecutor::ParallelExecutor(
     const std::unordered_set<std::string> &bcast_vars,
     const ProgramDesc &main_program, const std::string &loss_var_name,
     Scope *scope, const std::vector<Scope *> &local_scopes, bool allow_op_delay,
-    bool use_default_grad_scale)
+    bool use_default_grad_scale, bool balance_parameter_opt_between_cards)
     : member_(new ParallelExecutorPrivate(places)) {
   member_->global_scope_ = scope;
 
@@ -93,11 +93,12 @@ ParallelExecutor::ParallelExecutor(
 #ifdef PADDLE_WITH_CUDA
   details::MultiDevSSAGraphBuilder builder(
       member_->places_, loss_var_name, params, member_->local_scopes_,
-      member_->nccl_ctxs_.get(), use_default_grad_scale);
+      member_->nccl_ctxs_.get(), use_default_grad_scale,
+      balance_parameter_opt_between_cards);
 #else
-  details::MultiDevSSAGraphBuilder builder(member_->places_, loss_var_name,
-                                           params, member_->local_scopes_,
-                                           use_default_grad_scale);
+  details::MultiDevSSAGraphBuilder builder(
+      member_->places_, loss_var_name, params, member_->local_scopes_,
+      use_default_grad_scale, balance_parameter_opt_between_cards);
 #endif
   auto graph = builder.Build(main_program);
 
diff --git a/paddle/fluid/framework/parallel_executor.h b/paddle/fluid/framework/parallel_executor.h
@@ -40,7 +40,8 @@ class ParallelExecutor {
                             const ProgramDesc& main_program,
                             const std::string& loss_var_name, Scope* scope,
                             const std::vector<Scope*>& local_scopes,
-                            bool allow_op_delay, bool use_default_grad_scale);
+                            bool allow_op_delay, bool use_default_grad_scale,
+                            bool balance_parameter_opt_between_cards);
 
   ~ParallelExecutor();
 
diff --git a/paddle/fluid/pybind/pybind.cc b/paddle/fluid/pybind/pybind.cc
@@ -502,11 +502,13 @@ All parameter, weight, gradient are variables in Paddle.
               const std::unordered_set<std::string> &bcast_vars,
               const ProgramDesc &main_program, const std::string &loss_var_name,
               Scope *scope, std::vector<Scope *> &local_scopes,
-              bool allow_op_delay, bool use_default_grad_scale) {
+              bool allow_op_delay, bool use_default_grad_scale,
+              bool balance_parameter_opt_between_cards) {
              new (&self) ParallelExecutor(
                  num_threads, use_event, places, params, bcast_vars,
                  main_program, loss_var_name, scope, local_scopes,
-                 allow_op_delay, use_default_grad_scale);
+                 allow_op_delay, use_default_grad_scale,
+                 balance_parameter_opt_between_cards);
            })
       .def("bcast_params", &ParallelExecutor::BCastParamsToGPUs)
       // NOTE: even we return a vec<Scope*>* to Python use reference policy.
diff --git a/python/paddle/fluid/parallel_executor.py b/python/paddle/fluid/parallel_executor.py
@@ -30,7 +30,8 @@ def __init__(self,
                  num_threads=None,
                  allow_op_delay=False,
                  share_vars_from=None,
-                 use_default_grad_scale=True):
+                 use_default_grad_scale=True,
+                 balance_parameter_opt_between_cards=False):
         """
         ParallelExecutor can run program in parallel.
 
@@ -51,6 +52,9 @@ def __init__(self,
                 gradients of each device and scaled gradients would be
                 aggregated. Otherwise, a customized scale value should be fed
                 to the network.
+            balance_parameter_opt_between_cards(bool, default True): Whether
+                updating different gradients on different cards. Currently, it
+                is not recommended.
 
         Returns:
             A ParallelExecutor object.
@@ -129,7 +133,8 @@ def __init__(self,
             scope,
             local_scopes,
             allow_op_delay,
-            use_default_grad_scale)
+            use_default_grad_scale,
+            balance_parameter_opt_between_cards)
 
         self.scope = scope
 
diff --git a/python/paddle/fluid/tests/unittests/test_parallel_executor.py b/python/paddle/fluid/tests/unittests/test_parallel_executor.py
