[GPUPS]fix merge_grad&push_sparse (PaddlePaddle#25)

* fix merge_grad

* fix push_sparse

* fix push_sparse

* fix size_t

* change typo

paddle/fluid/framework/fleet/heter_ps/heter_comm_inl.h

@@ -118,8 +118,8 @@ __global__ void merge_gradient_kernel(const uint32_t* offset,
 
     for (int j = 1; j < num; ++j) {
       ori_index = index[start + j];
-      in = *(FeaturePushValue*)(input + size_t(ori_index) * grad_value_size);
-      merger_.add_basic_field(lhs, in);
+      FeaturePushValue& rhs = *(FeaturePushValue*)(input + size_t(ori_index) * grad_value_size);
+      merger_.add_basic_field(lhs, rhs);
     }
   }
 
@@ -954,7 +954,7 @@ void HeterComm<KeyType, ValType, GradType>::push_sparse(int gpu_num,
   auto stream = resource_->local_stream(gpu_num, 0);
 
   size_t grad_value_size =
-      TYPEALIGN(8, sizeof(FeaturePushValue) + (max_mf_dim_ * sizeof(float)));
+    TYPEALIGN(8, sizeof(FeaturePushValue) + (max_mf_dim_ * sizeof(float)));
 
   // int h_left[total_gpu];   // NOLINT
   // int h_right[total_gpu];  // NOLINT
@@ -976,14 +976,8 @@ void HeterComm<KeyType, ValType, GradType>::push_sparse(int gpu_num,
   auto d_shard_keys = memory::Alloc(place, len * sizeof(KeyType));
   KeyType* d_shard_keys_ptr = reinterpret_cast<KeyType*>(d_shard_keys->ptr());
 
-  GradType* d_shard_grads_ptr;
-  if (!multi_mf_dim_) {
-    auto d_shard_grads = memory::Alloc(place, len * sizeof(GradType));
-    d_shard_grads_ptr = reinterpret_cast<GradType*>(d_shard_grads->ptr());
-  } else {
-    auto d_shard_grads = memory::Alloc(place, len * grad_value_size);
-    d_shard_grads_ptr = reinterpret_cast<GradType*>(d_shard_grads->ptr());
-  }
+  auto d_shard_grads = memory::Alloc(place, len * grad_value_size);
+  GradType* d_shard_grads_ptr = reinterpret_cast<GradType*>(d_shard_grads->ptr());
 
   int uniq_len = len;
   merge_grad(gpu_num, d_keys, d_grads, NULL, len, uniq_len);