From e9d815e32b7cdb6e030bfd3aa649d3327bf4f195 Mon Sep 17 00:00:00 2001
From: typhoonzero <typhoonzero1986@gmail.com>
Date: Wed, 21 Mar 2018 14:46:10 +0800
Subject: [PATCH] prepare and create op before run

---
 paddle/fluid/operators/listen_and_serv_op.cc | 9 +--------
 paddle/fluid/operators/send_op.cc            | 1 +
 2 files changed, 2 insertions(+), 8 deletions(-)

diff --git a/paddle/fluid/operators/listen_and_serv_op.cc b/paddle/fluid/operators/listen_and_serv_op.cc
index bd6e25449f0..da44128cddd 100644
--- a/paddle/fluid/operators/listen_and_serv_op.cc
+++ b/paddle/fluid/operators/listen_and_serv_op.cc
@@ -99,7 +99,6 @@ class ListenAndServOp : public framework::OperatorBase {
     blk_ctx_list.push_back(nullptr);  // block0 is not used.
     for (int blkid = 1; blkid < num_blocks; ++blkid) {
       auto *exe_ctx = executor.Prepare(*program, blkid);
-      VLOG(2) << "prepare ctx: " << exe_ctx;
       blk_ctx_list.push_back(exe_ctx);
     }
 
@@ -149,6 +148,7 @@ class ListenAndServOp : public framework::OperatorBase {
       // should be global ops.
       // NOTE: if is_gpu_place, CUDA kernels are laugched by multiple threads
       // and this will still work.
+
       std::vector<std::future<void>> fs;
       // block0 contains only listen_and_serv op, start run from block1.
       for (int blkid = 1; blkid < num_blocks - 1; ++blkid) {
@@ -156,13 +156,8 @@ class ListenAndServOp : public framework::OperatorBase {
             [&executor, &program, &recv_scope, &blk_ctx_list, blkid]() {
               int run_block = blkid;  // thread local
               try {
-                VLOG(2) << "run ctx: " << blk_ctx_list[run_block]
-                        << " block: " << run_block;
                 executor.RunPreparedContext(blk_ctx_list[run_block],
                                             &recv_scope, false, false);
-                // executor.Run(*program, &recv_scope, run_block,
-                //              false /*create_local_scope*/,
-                //              false /*create_vars*/);
               } catch (std::exception &e) {
                 LOG(ERROR) << "run sub program error " << e.what();
               }
@@ -174,8 +169,6 @@ class ListenAndServOp : public framework::OperatorBase {
         try {
           executor.RunPreparedContext(blk_ctx_list[num_blocks - 1], &recv_scope,
                                       false, false);
-          // executor.Run(*program, &recv_scope, num_blocks - 1,
-          //              false /*create_local_scope*/, false /*create_vars*/);
         } catch (std::exception &e) {
           LOG(ERROR) << "run sub program error " << e.what();
         }
diff --git a/paddle/fluid/operators/send_op.cc b/paddle/fluid/operators/send_op.cc
index 443f40e803e..2df25ae5a6d 100644
--- a/paddle/fluid/operators/send_op.cc
+++ b/paddle/fluid/operators/send_op.cc
@@ -66,6 +66,7 @@ class SendOp : public framework::OperatorBase {
     auto* client_var = scope.FindVar(client_var_name);
     detail::RPCClient* rpc_client = client_var->GetMutable<detail::RPCClient>();
 
+    ctx.Wait();  // wait before sending
     for (size_t i = 0; i < ins.size(); i++) {
       if (NeedSend(scope, ins[i])) {
         VLOG(3) << "sending " << ins[i] << " to " << epmap[i];
-- 
GitLab