From 527946df490df1ad80152ffdc973178b9ae308f6 Mon Sep 17 00:00:00 2001
From: tangwei12 <tangwei12@baidu.com>
Date: Fri, 7 Dec 2018 18:08:29 +0800
Subject: [PATCH] add scope in prefetch

---
 .../distributed/parameter_prefetch.cc         | 19 +++++++-------
 paddle/fluid/operators/lookup_table_op.h      |  3 ++-
 paddle/fluid/operators/nce_op.h               | 25 ++++++++++++++-----
 3 files changed, 31 insertions(+), 16 deletions(-)

diff --git a/paddle/fluid/operators/distributed/parameter_prefetch.cc b/paddle/fluid/operators/distributed/parameter_prefetch.cc
index cf14538b1..67b56bd21 100644
--- a/paddle/fluid/operators/distributed/parameter_prefetch.cc
+++ b/paddle/fluid/operators/distributed/parameter_prefetch.cc
@@ -102,8 +102,9 @@ static void MergeMultipleVarsIntoOneBySection(
     const std::string& out_name, const std::vector<std::string>& out_var_names,
     const std::vector<int>& height_section,
     const std::vector<std::vector<int64_t>>& splited_ids,
-    const framework::ExecutionContext& context, framework::Scope* scope,
-    platform::DeviceContext* actual_ctx) {
+    const framework::ExecutionContext& context,
+    const framework::Scope& actual_scope, framework::Scope* scope,
+    platform::DeviceContext* actual_ctx, ) {
   PADDLE_ENFORCE_EQ(out_var_names.size(), height_section.size(), "");
 
   auto cpu_place = platform::CPUPlace();
@@ -114,9 +115,9 @@ static void MergeMultipleVarsIntoOneBySection(
     id_to_offset[ids_vector[i]].push_back(i);
   }
 
-  auto& id_tensor = scope->FindVar(id_name)->Get<framework::LoDTensor>();
+  auto& id_tensor = actual_scope.FindVar(id_name)->Get<framework::LoDTensor>();
   auto* out_tensor =
-      scope->FindVar(out_name)->GetMutable<framework::LoDTensor>();
+      actual_scope.FindVar(out_name)->GetMutable<framework::LoDTensor>();
   auto* out_tensor_data = out_tensor->mutable_data<float>(id_tensor.place());
 
   bool is_on_cpu_place = true;
@@ -172,8 +173,9 @@ void prefetch(const std::string& id_name, const std::string& out_name,
               const std::vector<std::string>& table_names,
               const std::vector<std::string>& epmap,
               const std::vector<int>& height_sections,
-              const framework::ExecutionContext& context) {
-  auto& local_scope = context.scope().NewScope();
+              const framework::ExecutionContext& context,
+              const framework::Scope& scope) {
+  auto& local_scope = scope.NewScope();
 
   platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance();
   auto& cpu_ctx = *pool.Get(platform::CPUPlace());
@@ -245,9 +247,8 @@ void prefetch(const std::string& id_name, const std::string& out_name,
 
   MergeMultipleVarsIntoOneBySection(id_name, ids_vector, out_name,
                                     out_var_names, height_sections, splited_ids,
-                                    context, &local_scope, &actual_ctx);
-
-  context.scope().DeleteScope(&local_scope);
+                                    context, scope, &local_scope, &actual_ctx);
+  scope.DeleteScope(&local_scope);
 }
 
 };  // namespace distributed
diff --git a/paddle/fluid/operators/lookup_table_op.h b/paddle/fluid/operators/lookup_table_op.h
index 3a73a7637..a7d0fd485 100644
--- a/paddle/fluid/operators/lookup_table_op.h
+++ b/paddle/fluid/operators/lookup_table_op.h
@@ -59,7 +59,8 @@ class LookupTableKernel : public framework::OpKernel<T> {
 // server
 #ifdef PADDLE_WITH_DISTRIBUTE
       operators::distributed::prefetch(id_name, out_name, table_names, epmap,
-                                       height_sections, context);
+                                       height_sections, context,
+                                       context.scope());
 #else
       PADDLE_THROW(
           "paddle is not compiled with distribute support, can not do "
diff --git a/paddle/fluid/operators/nce_op.h b/paddle/fluid/operators/nce_op.h
index 6567b6534..9789e3038 100644
--- a/paddle/fluid/operators/nce_op.h
+++ b/paddle/fluid/operators/nce_op.h
@@ -170,18 +170,31 @@ class NCEKernel : public framework::OpKernel<T> {
       auto height_sections = context.Attr<std::vector<int>>("height_sections");
       auto table_names = context.Attr<std::vector<std::string>>("table_names");
 
-      local_scope.Var("Ids");
-      local_scope.Var("Weight");
+      auto *ids = local_scope.Var("Ids");
+      auto *x_tensor = ids->GetMutable<framework::LoDTensor>();
+      x_tensor->mutable_data<int64_t>(
+          framework::make_ddim({static_cast<int64_t>(labels.size()), 1}),
+          context.GetPlace());
+      // copy.
+      std::memcpy(x_tensor->data<int64_t>(), labels.data(),
+                  labels.size() * sizeof(int64_t));
+
+      local_scope.Var("Weight@Local")
+          ->GetMutable<framework::Tensor>()
+          ->mutable_data<T>(context.GetPlace());
 
 #ifdef PADDLE_WITH_DISTRIBUTE
-      operators::distributed::prefetch("Ids", "Weight", table_names, epmap,
-                                       height_sections, context);
+      operators::distributed::prefetch("Ids", "Weight@Local", table_names,
+                                       epmap, height_sections, context,
+                                       &local_scope);
 #else
       PADDLE_THROW(
           "paddle is not compiled with distribute support, can not do "
           "parameter prefetch!");
+#endif
 
-      auto weight_mat = EigenMatrix<T>::From(*(weight->Get<T>()));
+      auto weight_mat = EigenMatrix<T>::From(
+          (local_scope.Var("Weight@Local")->Get<framework::Tensor>()));
       for (int64_t i = 0; i < sample_labels->numel(); ++i) {
         std::vector<int64_t>::iterator it =
             std::find(labels.begin(), labels.end(), sample_labels_data[i]);
@@ -196,7 +209,7 @@ class NCEKernel : public framework::OpKernel<T> {
       }
 
       context.scope().DeleteScope(&local_scope);
-#endif
+
     } else {
       auto weight_mat =
           EigenMatrix<T>::From(*(context.Input<Tensor>("Weight")));
-- 
GitLab