"fix unique_ptr copy"

Author: dzhwinter

paddle/operators/merge_lod_tensor_op.cc

@@ -45,7 +45,7 @@ class MergeLoDTensorOp : public framework::OperatorBase {
       cpu_mask->ShareDataWith(mask);
     } else if (platform::is_gpu_place(mask.place())) {
 #ifdef PADDLE_WITH_CUDA
-      framework::CopyFrom(mask, platform::CPUPlace(), dev_ctx, &cpu_mask);
+      framework::CopyFrom(mask, platform::CPUPlace(), dev_ctx, cpu_mask.get());
 #else
       PADDLE_THROW("Not supported GPU, Please compile WITH_GPU option");
 #endif

paddle/operators/multiplex_op.cu

@@ -33,7 +33,7 @@ class MultiplexGPUKernel : public framework::OpKernel<T> {
     auto cols = ins[0]->numel() / rows;
     // copy index to cpu
     Tensor index_t_cpu;
-    index_t_cpu.CopyFrom(*ids, platform::CPUPlace(), ctx.device_context());
+    CopyFrom(*ids, platform::CPUPlace(), ctx.device_context(), &index_t_cpu);
     auto* index = index_t_cpu.data<int32_t>();
     auto stream = ctx.cuda_device_context().stream();
     Place place = boost::get<Place>(ctx.GetPlace());
@@ -68,7 +68,7 @@ class MultiplexGradGPUKernel : public framework::OpKernel<T> {
     auto cols = ins[0]->numel() / rows;
     // copy index to cpu
     Tensor index_t_cpu;
-    index_t_cpu.CopyFrom(*ids, platform::CPUPlace(), ctx.device_context());
+    CopyFrom(*ids, platform::CPUPlace(), ctx.device_context(), &index_t_cpu);
     auto* index = index_t_cpu.data<int32_t>();
 
     auto stream = ctx.cuda_device_context().stream();

paddle/operators/split_lod_tensor_op.cc

@@ -49,7 +49,7 @@ class SplitLoDTensorOp : public framework::OperatorBase {
       cpu_mask->ShareDataWith(mask);
     } else if (platform::is_gpu_place(mask.place())) {
 #ifdef PADDLE_WITH_CUDA
-      framework::CopyFrom(mask, platform::CPUPlace(), dev_ctx, &cpu_mask);
+      framework::CopyFrom(mask, platform::CPUPlace(), dev_ctx, cpu_mask.get());
 #else
       PADDLE_THROW("Not supported GPU, Please compile WITH_GPU option");
 #endif