remove exe context

paddle/fluid/operators/distributed/parameter_recv.cc

@@ -40,7 +40,6 @@ using DDim = framework::DDim;
 
 template <typename T>
 void ParameterRecv<T>::operator()(const RpcContext &rpc_ctx,
-                                  const framework::ExecutionContext &ctx,
                                   const framework::Scope &scope) {
   framework::Scope *local_scope = scope.NewTmpScope();
 
@@ -48,8 +47,7 @@ void ParameterRecv<T>::operator()(const RpcContext &rpc_ctx,
   auto &cpu_ctx = *pool.Get(platform::CPUPlace());
 
   distributed::RPCClient *rpc_client =
-      distributed::RPCClient::GetInstance<RPCCLIENT_T>(
-          ctx.Attr<int>("trainer_id"));
+      distributed::RPCClient::GetInstance<RPCCLIENT_T>(0);
 
   auto *recv_var = scope.FindVar(rpc_ctx.var_name);
 
@@ -80,12 +78,13 @@ void ParameterRecv<T>::operator()(const RpcContext &rpc_ctx,
     size_t output_offset = 0;
     framework::Tensor *recv_tensor =
         recv_var->GetMutable<framework::LoDTensor>();
+    auto dev_ctx = paddle::platform::CPUDeviceContext();
     for (auto *in : recved_tensors) {
       auto in_stride = framework::stride_numel(in->dims());
       auto out_stride = framework::stride_numel(recv_tensor->dims());
       StridedNumelCopyWithAxis<T>(
-          ctx.device_context(), 0, recv_tensor->data<T>() + output_offset,
-          out_stride, in->data<T>(), in_stride, in_stride[0]);
+          dev_ctx, 0, recv_tensor->data<T>() + output_offset, out_stride,
+          in->data<T>(), in_stride, in_stride[0]);
       output_offset += in_stride[0];
     }
   }

paddle/fluid/operators/distributed/parameter_recv.h

@@ -26,9 +26,7 @@ namespace distributed {
 
 template <typename T>
 struct ParameterRecv {
-  void operator()(const RpcContext &rpc_ctx,
-                  const framework::ExecutionContext &context,
-                  const framework::Scope &scope);
+  void operator()(const RpcContext &rpc_ctx, const framework::Scope &scope);
 };
 
 };  // namespace distributed

paddle/fluid/operators/distributed/parameter_send.cc

@@ -39,7 +39,6 @@ using DDim = framework::DDim;
 
 template <typename T>
 void ParameterSend<T>::operator()(const RpcContext &rpc_ctx,
-                                  const framework::ExecutionContext &ctx,
                                   const framework::Scope &scope, bool sync) {
   framework::Scope *local_scope = scope.NewTmpScope();
 
@@ -47,8 +46,7 @@ void ParameterSend<T>::operator()(const RpcContext &rpc_ctx,
   auto &cpu_ctx = *pool.Get(platform::CPUPlace());
 
   distributed::RPCClient *rpc_client =
-      distributed::RPCClient::GetInstance<RPCCLIENT_T>(
-          ctx.Attr<int>("trainer_id"));
+      distributed::RPCClient::GetInstance<RPCCLIENT_T>(0);
 
   auto *send_var = scope.FindVar(rpc_ctx.var_name);
   size_t out_num = rpc_ctx.splited_var_names.size();
@@ -105,7 +103,7 @@ void ParameterSend<T>::operator()(const RpcContext &rpc_ctx,
       outs_rows_idx[out_idx].push_back(send_rows[i]);
       outs_dense_idx[out_idx].push_back(i);
     }
-    auto place = ctx.GetPlace();
+    auto place = platform::CPUPlace();
 
     for (size_t i = 0; i < outs_rows_idx.size(); ++i) {
       auto rows_idx = outs_rows_idx[i];
@@ -118,22 +116,25 @@ void ParameterSend<T>::operator()(const RpcContext &rpc_ctx,
         for (auto idx : rows_idx) {
           outs[i]->mutable_rows()->push_back(idx - abs_sections[i]);
         }
-        auto dst = outs[i]->mutable_value()->mutable_data<T>(ctx.GetPlace());
+        auto dst = outs[i]->mutable_value()->mutable_data<T>(place);
         for (size_t j = 0; j < rows_idx.size(); j++) {
           if (platform::is_cpu_place(place)) {
             memory::Copy(
                 platform::CPUPlace(), dst + j * row_numel, platform::CPUPlace(),
                 src + outs_dense_idx[i][j] * row_numel, sizeof(T) * row_numel);
           } else {
-#ifdef PADDLE_WITH_CUDA
+            PADDLE_THROW("do not support GPU now");
+            /*
+            #ifdef PADDLE_WITH_CUDA
                         auto stream = ctx.cuda_device_context().stream();
                         memory::Copy(platform::CUDAPlace(), dst + j * row_numel,
                                      platform::CUDAPlace(),
                                      src + outs_dense_idx[i][j] * row_numel,
                                      sizeof(T) * row_numel, stream);
-#else
+            #else
                         PADDLE_THROW("Paddle is not compiled with GPU");
-#endif
+            #endif
+            */
           }
         }
       }

paddle/fluid/operators/distributed/parameter_send.h

@@ -26,9 +26,8 @@ namespace distributed {
 
 template <typename T>
 struct ParameterSend {
-  void operator()(const RpcContext &rpc_ctx,
-                  const framework::ExecutionContext &context,
-                  const framework::Scope &scope, bool sync);
+  void operator()(const RpcContext &rpc_ctx, const framework::Scope &scope,
+                  bool sync);
 };
 
 };  // namespace distributed

paddle/fluid/operators/distributed_ops/recv_op.cc

@@ -62,7 +62,7 @@ class RecvOp : public framework::OperatorBase {
           framework::ExecutionContext(*this, scope, *dev_ctx, ctx, nullptr);
       auto recv_functor = distributed::ParameterRecv<float>();
       auto rpc_ctx = distributed::RpcContext(outs[0], recv_varnames, epmap, {});
-      recv_functor(rpc_ctx, exe_ctx, scope);
+      recv_functor(rpc_ctx, scope);
     } else {
       if (with_barrier) {
         std::vector<distributed::VarHandlePtr> rets;

paddle/fluid/operators/distributed_ops/send_op.cc

@@ -56,7 +56,7 @@ class SendOp : public framework::OperatorBase {
       auto send_functor = distributed::ParameterSend<float>();
       auto rpc_ctx = distributed::RpcContext(ins[0], send_varnames, epmap,
                                              height_sections);
-      send_functor(rpc_ctx, exe_ctx, scope, static_cast<bool>(sync_send));
+      send_functor(rpc_ctx, scope, static_cast<bool>(sync_send));
     } else {
       platform::DeviceContextPool& pool =
           platform::DeviceContextPool::Instance();