Reverse iterator op's input

test=develop

paddle/fluid/imperative/layer.cc

@@ -81,10 +81,6 @@ class TensorAddToFunctor : public boost::static_visitor<> {
 
 }  // namespace detail
 
-template <int MajorType = Eigen::RowMajor,
-          typename IndexType = Eigen::DenseIndex>
-using EigenVector = framework::EigenVector<float, MajorType, IndexType>;
-
 void AddTo(Variable* src, Variable* dst, platform::Place place) {
   framework::Tensor* dst_tensor = dst->GetMutable<framework::LoDTensor>();
   framework::Tensor* src_tensor = src->GetMutable<framework::LoDTensor>();
@@ -99,18 +95,10 @@ void AddTo(Variable* src, Variable* dst, platform::Place place) {
                  "dst_numel %lld vs. src_numel %lld", dst_tensor->numel(),
                  src_tensor->numel());
 
-  auto result = EigenVector<>::Flatten(*dst_tensor);
-  auto in_0_e = EigenVector<>::Flatten(*dst_tensor);
-  auto in_1_e = EigenVector<>::Flatten(*src_tensor);
-  platform::DeviceContext* dev_ctx =
-      platform::DeviceContextPool::Instance().Get(place);
-  platform::CPUDeviceContext* x =
-      reinterpret_cast<platform::CPUDeviceContext*>(dev_ctx);
-  result.device(*x->eigen_device()) = in_0_e + in_1_e;
-  // detail::TensorAddToFunctor<float> func(
-  // src_tensor->numel(), src_tensor->data<float>(),
-  // dst_tensor->mutable_data<float>(place));
-  // boost::apply_visitor(func, place);
+  detail::TensorAddToFunctor<float> func(
+      src_tensor->numel(), src_tensor->data<float>(),
+      dst_tensor->mutable_data<float>(place));
+  boost::apply_visitor(func, place);
 }
 
 class Autograd {
@@ -134,7 +122,7 @@ class Autograd {
       std::map<std::string, std::vector<VarBase*>> input_grads =
           ready_op->ApplyGrad();
 
-      for (auto it : input_grads) {
+      for (auto it = input_grads.rbegin(); it != input_grads.rend(); ++it) {
         const std::vector<VarBase*>& ingrads = it.second;
         for (int64_t i = ingrads.size() - 1; i >= 0; --i) {
           if (!ingrads[i]) continue;

python/paddle/fluid/framework.py

@@ -104,14 +104,14 @@ def cuda_places(device_ids=None):
     :code:`FLAGS_selected_gpus=0,1,2`, the returned list would
     be [fluid.CUDAPlace(0), fluid.CUDAPlace(1), fluid.CUDAPlace(2)].
     If :code:`FLAGS_selected_gpus` is not set, all visible
-    gpu places would be returned.
+    gpu places would be returned.  
 
     If :code:`device_ids` is not None, it should be the device
-    ids of gpus. For example, if :code:`device_ids=[0,1,2]`,
-    the returned list would be
+    ids of gpus. For example, if :code:`device_ids=[0,1,2]`, 
+    the returned list would be 
     [fluid.CUDAPlace(0), fluid.CUDAPlace(1), fluid.CUDAPlace(2)].
-
-    Args:
+    
+    Args: 
         device_ids (None|list(int)|tuple(int)): gpu device id list.
 
     Returns:
@@ -133,11 +133,11 @@ def cuda_places(device_ids=None):
 def cpu_places(device_count=None):
     '''
     Create a list of :code:`fluid.CPUPlace` objects.
-
+    
     If :code:`device_count` is None, the device count would
-    be determined by environment variable :code:`CPU_NUM`.
+    be determined by environment variable :code:`CPU_NUM`. 
     If :code:`CPU_NUM` is not set, the device count would
-    be determined by :code:`multiprocessing.cpu_count()`.
+    be determined by :code:`multiprocessing.cpu_count()`. 
 
     Args:
         device_count (None|int): device number.
@@ -155,9 +155,9 @@ def cuda_pinned_places(device_count=None):
     Create a list of :code:`fluid.CUDAPinnedPlace` objects.
 
     If :code:`device_count` is None, the device count would
-    be determined by environment variable :code:`CPU_NUM`.
+    be determined by environment variable :code:`CPU_NUM`. 
     If :code:`CPU_NUM` is not set, the device count would
-    be determined by :code:`multiprocessing.cpu_count()`.
+    be determined by :code:`multiprocessing.cpu_count()`. 
 
     Args:
         device_count (None|int): device number.
@@ -2716,6 +2716,11 @@ class Program(object):
         # whether the program is optimized by memory_optimize_transpiler
         self.__is_mem_optimized = False
 
+        # if this program has been optimized by distributed optimizer
+        # fleet_opt will be given a value
+        self._fleet_opt = None
+        self._program_config = None
+
     @property
     def _is_mem_optimized(self):
         # if the program is optimized, operator input/outputs