From 179b78934a81c7935b3a3d6fa22f9596170a31dc Mon Sep 17 00:00:00 2001
From: dzhwinter <dzhwinter@gmail.com>
Date: Tue, 6 Feb 2018 00:24:13 -0800
Subject: [PATCH] "fix CopyToPeer"

---
 paddle/framework/lod_tensor.h      |  2 +-
 paddle/framework/mixed_vector.h    | 25 +++++++++++++++++++++++--
 paddle/operators/parallel_do_op.cc |  4 ++--
 3 files changed, 26 insertions(+), 5 deletions(-)

diff --git a/paddle/framework/lod_tensor.h b/paddle/framework/lod_tensor.h
index a773c1eb32d..be2b3016196 100644
--- a/paddle/framework/lod_tensor.h
+++ b/paddle/framework/lod_tensor.h
@@ -65,7 +65,7 @@ struct LoD : public std::vector<Vector<size_t>> {
 
   void CopyToPeer(platform::Place place) {
     for (auto it = this->begin(); it != this->end(); ++it) {
-      it->mutable_data(place);
+      it->CopyToPeer(place);
     }
   }
 };
diff --git a/paddle/framework/mixed_vector.h b/paddle/framework/mixed_vector.h
index 1fc7622e9b2..cdb968e3cb7 100644
--- a/paddle/framework/mixed_vector.h
+++ b/paddle/framework/mixed_vector.h
@@ -82,7 +82,7 @@ inline const T *Vector<T>::data(platform::Place place) const {
     if (cuda_ptr_ == nullptr) {
       return nullptr;
     }
-    if (platform::is_same_place(place, place_)) {
+    if (boost::get<platform::CUDAPlace>(place) == place_) {
       return static_cast<const T *>(cuda_ptr_.get());
     } else {
       PADDLE_THROW(
@@ -99,7 +99,7 @@ inline T *Vector<T>::mutable_data(platform::Place place) {
   if (platform::is_cpu_place(place)) {
     return std::vector<T>::data();
   } else if (platform::is_gpu_place(place)) {
-    if (!platform::is_same_place(place, place_)) {
+    if (boost::get<platform::CUDAPlace>(place) != place_) {
       place_ = boost::get<platform::CUDAPlace>(place);
     }
 #ifdef PADDLE_WITH_CUDA
@@ -159,5 +159,26 @@ void Vector<T>::CopyFromCUDA() {
 #endif
 }
 
+template <typename T>
+void Vector<T>::CopyToPeer(platform::Place place) {
+#ifdef PADDLE_WITH_CUDA
+  if (boost::get<platform::CUDAPlace>(place) != place_) {
+    place_ = boost::get<platform::CUDAPlace>(place);
+  }
+  if (cuda_size_ < this->size() || cuda_ptr_ == nullptr) {
+    cuda_ptr_.reset(
+        memory::Alloc<platform::CUDAPlace>(place_, this->size() * sizeof(T)),
+        memory::PlainDeleter<void, platform::CUDAPlace>(place_));
+  }
+  cuda_size_ = this->size();
+  platform::DeviceContextPool &pool = platform::DeviceContextPool::Instance();
+  auto *ctx = pool.GetByPlace(place_);
+  memory::Copy(place_, cuda_ptr_.get(), platform::CPUPlace(),
+               static_cast<const void *>(this->data()),
+               this->size() * sizeof(T), ctx->stream());
+  ctx->Wait();
+#endif
+}
+
 }  // namespace framework
 }  // namespace paddle
diff --git a/paddle/operators/parallel_do_op.cc b/paddle/operators/parallel_do_op.cc
index 0db2fb6238a..eb6308d306a 100644
--- a/paddle/operators/parallel_do_op.cc
+++ b/paddle/operators/parallel_do_op.cc
@@ -79,7 +79,7 @@ inline void CopyOrShare(const framework::Variable &src,
       dst->GetMutable<LoDTensor>()->set_lod(src.Get<LoDTensor>().lod());
     } else {
       Copy(src.Get<LoDTensor>(), dst_place, dst->GetMutable<LoDTensor>());
-      LoD lod(src.Get<LoDTensor>().lod());
+      framework::LoD lod(src.Get<LoDTensor>().lod());
       lod.CopyToPeer(dst_place);
       dst->GetMutable<LoDTensor>()->set_lod(lod);
     }
@@ -92,7 +92,7 @@ inline void CopyOrShare(const framework::Variable &src,
       dst_sr->set_rows(src_sr.rows());
     } else {
       Copy(src_sr.value(), dst_place, dst_sr->mutable_value());
-      LoD lod(src.Get<LoDTensor>().lod());
+      framework::Vector<int64_t> lod(src_sr.rows());
       lod.CopyToPeer(dst_place);
       dst_sr->set_rows(lod);
     }
-- 
GitLab