From e72637ddd22765dd915119b96bc1821734cd28ef Mon Sep 17 00:00:00 2001
From: Qiao Longfei <qiaolongfei@baidu.com>
Date: Sat, 9 Feb 2019 17:11:46 +0800
Subject: [PATCH] ThreadedSSAGraphExecutor support num_iteration_per_run
 test=develop

---
 .../details/async_ssa_graph_executor.cc       | 16 ------------
 .../details/threaded_ssa_graph_executor.cc    | 25 +++++++++++++++++--
 .../details/threaded_ssa_graph_executor.h     |  1 +
 3 files changed, 24 insertions(+), 18 deletions(-)

diff --git a/paddle/fluid/framework/details/async_ssa_graph_executor.cc b/paddle/fluid/framework/details/async_ssa_graph_executor.cc
index 79b390dde48..5ce92ad8267 100644
--- a/paddle/fluid/framework/details/async_ssa_graph_executor.cc
+++ b/paddle/fluid/framework/details/async_ssa_graph_executor.cc
@@ -30,19 +30,6 @@ AsyncSSAGraphExecutor::AsyncSSAGraphExecutor(
   VLOG(3) << "build AsyncSSAGraphExecutor";
   PADDLE_ENFORCE_EQ(places_.size(), local_scopes_.size());
 
-  if (strategy_.num_iteration_per_run_ > 1) {
-    int read_op_num = 0;
-    for (auto *node : graphs_[0]->Nodes()) {
-      if (node->IsOp() && node->Name() == "read") {
-        read_op_num++;
-      }
-    }
-    if (read_op_num == 0) {
-      LOG(WARNING) << "when num_iteration_per_run_ is larger then 1, the model "
-                      "should use pyreader to feed data!";
-    }
-  }
-
   // set the correct size of thread pool to each device.
   strategy_.num_threads_ = strategy_.num_threads_ < places_.size()
                                ? 1UL
@@ -69,9 +56,6 @@ FeedFetchList AsyncSSAGraphExecutor::Run(
   for (size_t i = 0; i < places_.size(); ++i) {
     auto call = [this, i, &fetch_tensors]() -> FeedFetchList {
       try {
-        for (size_t j = 0; j < strategy_.num_iteration_per_run_ - 1; ++j) {
-          executors_[i]->Run(fetch_tensors);
-        }
         return executors_[i]->Run(fetch_tensors);
       } catch (...) {
         exception_holder_.Catch(std::current_exception());
diff --git a/paddle/fluid/framework/details/threaded_ssa_graph_executor.cc b/paddle/fluid/framework/details/threaded_ssa_graph_executor.cc
index 677a2937945..16fa2a6db68 100644
--- a/paddle/fluid/framework/details/threaded_ssa_graph_executor.cc
+++ b/paddle/fluid/framework/details/threaded_ssa_graph_executor.cc
@@ -32,9 +32,22 @@ ThreadedSSAGraphExecutor::ThreadedSSAGraphExecutor(
       places_(places),
       fetch_ctxs_(places),
       running_ops_(0),
-      strategy_(strategy) {}
+      strategy_(strategy) {
+  if (strategy_.num_iteration_per_run_ > 1) {
+    int read_op_num = 0;
+    for (auto *node : graph_->Nodes()) {
+      if (node->IsOp() && node->Name() == "read") {
+        read_op_num++;
+      }
+    }
+    if (read_op_num == 0) {
+      LOG(WARNING) << "when num_iteration_per_run_ is larger then 1, the model "
+                      "should use pyreader to feed data!";
+    }
+  }
+}
 
-FeedFetchList ThreadedSSAGraphExecutor::Run(
+inline FeedFetchList ThreadedSSAGraphExecutor::RunImpl(
     const std::vector<std::string> &fetch_tensors) {
   std::unique_ptr<platform::RecordEvent> event(
       new platform::RecordEvent("ThreadedSSAGraphExecutorPrepare", nullptr));
@@ -140,6 +153,14 @@ FeedFetchList ThreadedSSAGraphExecutor::Run(
   return fetch_data;
 }
 
+FeedFetchList ThreadedSSAGraphExecutor::Run(
+    const std::vector<std::string> &fetch_tensors) {
+  for (size_t j = 0; j < strategy_.num_iteration_per_run_ - 1; ++j) {
+    RunImpl({});
+  }
+  return RunImpl(fetch_tensors);
+}
+
 void ThreadedSSAGraphExecutor::InsertFetchOps(
     const std::vector<std::string> &fetch_tensors,
     std::vector<FetchOpHandle *> *fetch_ops,
diff --git a/paddle/fluid/framework/details/threaded_ssa_graph_executor.h b/paddle/fluid/framework/details/threaded_ssa_graph_executor.h
index 24da56c09e3..3809b6e9ae0 100644
--- a/paddle/fluid/framework/details/threaded_ssa_graph_executor.h
+++ b/paddle/fluid/framework/details/threaded_ssa_graph_executor.h
@@ -51,6 +51,7 @@ class ThreadedSSAGraphExecutor : public SSAGraphExecutor {
   ~ThreadedSSAGraphExecutor() final = default;
 
  private:
+  inline FeedFetchList RunImpl(const std::vector<std::string> &fetch_tensors);
   void RunOp(const std::shared_ptr<BlockingQueue<VarHandleBase *>> &ready_var_q,
              details::OpHandleBase *op);
 
-- 
GitLab