Resovle multi gpu async deps (#12828)

* dist transpiler add control dependency var between send and recv

* fix async deps

* follow comments and refine

* fix deps connect for rpc ops

paddle/fluid/framework/details/multi_devices_graph_pass.cc

@@ -763,6 +763,8 @@ void MultiDevSSAGraphBuilder::CreateDistTrainOp(ir::Graph *result,
 // Create RPC related op handles that connects its in ops and out ops.
 void MultiDevSSAGraphBuilder::CreateRPCOp(ir::Graph *result,
                                           ir::Node *node) const {
+  // FIXME(typhoonzero): Cleanup this deps for both sync mode and async mode
+  //                     put them into transpiler.
   int op_dev_id = -1;
   if (node->Op()->Type() == "send") {
     // TODO(paddle-dev): getting the first var is not safe.
@@ -771,26 +773,42 @@ void MultiDevSSAGraphBuilder::CreateRPCOp(ir::Graph *result,
                    "This hack no longer holds, please fix.");
     // the variable name which contains .block means it was splited by
     // split_byref op
-    // so that we can balance the variable blocks to all the pserver
-    // instances.
     if (strategy_.reduce_ == BuildStrategy::ReduceStrategy::kAllReduce &&
         node->inputs[0]->Name().find(".block") == std::string::npos) {
       std::vector<std::string> input_var_names;
       for (ir::Node *n : node->inputs) {
         input_var_names.push_back(n->Name());
       }
-      op_dev_id = GetAppropriateDeviceID(input_var_names);
+      auto send_param_grad = boost::get<std::vector<std::string>>(
+          node->Op()->GetAttr(OpProtoAndCheckerMaker::OpRoleVarAttrName()));
+      PADDLE_ENFORCE_EQ(send_param_grad.size(), 2U);
+      op_dev_id = GetAppropriateDeviceID({send_param_grad[1]});
+      VLOG(10) << "send grad " << input_var_names[0] << " origin "
+               << send_param_grad[1] << " place: " << op_dev_id;
       for (auto &varname : input_var_names) {
         result->Get<ShardedVarDevice>(kShardedVarDevice)
             .emplace(varname, op_dev_id);
       }
+      result->Get<ShardedVarDevice>(kShardedVarDevice)
+          .emplace(send_param_grad[1], op_dev_id);
     }
   } else if (node->Op()->Type() == "recv") {
     std::vector<std::string> output_var_names;
     for (ir::Node *n : node->outputs) {
       output_var_names.push_back(n->Name());
     }
-    op_dev_id = GetAppropriateDeviceID(output_var_names);
+    auto recv_param_grad = boost::get<std::vector<std::string>>(
+        node->Op()->GetAttr(OpProtoAndCheckerMaker::OpRoleVarAttrName()));
+    // FIXME(typhoonzero): assume each recv op output one param
+    // Use the same place as send.
+    if (recv_param_grad.size() == 2U) {
+      op_dev_id = GetVarDeviceID(*result, recv_param_grad[1]);
+      VLOG(10) << "recv param " << recv_param_grad[0]
+               << " get grad place: " << recv_param_grad[1]
+               << " place: " << op_dev_id;
+    } else {
+      op_dev_id = GetAppropriateDeviceID(output_var_names);
+    }
     for (auto &varname : output_var_names) {
       result->Get<ShardedVarDevice>(kShardedVarDevice)
           .emplace(varname, op_dev_id);

paddle/fluid/framework/ir/node.cc

@@ -17,7 +17,7 @@ limitations under the License. */
 namespace paddle {
 namespace framework {
 namespace ir {
-const char Node::kControlDepVarName[] = "__control_var";
+constexpr char Node::kControlDepVarName[];
 }  // namespace ir
 }  // namespace framework
 }  // namespace paddle

paddle/fluid/framework/ir/node.h

@@ -27,7 +27,7 @@ namespace ir {
 class Node {
  public:
   enum class Type { kOperation, kVariable };
-  static const char kControlDepVarName[];
+  static constexpr char kControlDepVarName[] = "__control_var";
 
   explicit Node(const std::string& name, Type type)
       : name_(name), var_desc_(nullptr), op_desc_(nullptr), type_(type) {}

paddle/fluid/pybind/const_value.cc

@@ -13,7 +13,8 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #include "paddle/fluid/pybind/const_value.h"
-#include <paddle/fluid/framework/op_proto_maker.h>
+#include "paddle/fluid/framework/ir/node.h"
+#include "paddle/fluid/framework/op_proto_maker.h"
 #include "paddle/fluid/framework/operator.h"
 
 namespace paddle {
@@ -24,6 +25,8 @@ void BindConstValue(pybind11::module* m) {
   m->def("kTempVarName", [] { return framework::kTempVarName; });
   m->def("kGradVarSuffix", [] { return framework::kGradVarSuffix; });
   m->def("kZeroVarSuffix", [] { return framework::kZeroVarSuffix; });
+  m->def("kControlDepVarName",
+         [] { return framework::ir::Node::kControlDepVarName; });
 
   auto op_proto_and_checker_maker =
       m->def_submodule("op_proto_and_checker_maker");

python/paddle/fluid/framework.py

@@ -49,6 +49,12 @@ EMPTY_VAR_NAME = core.kEmptyVarName()
 TEMP_VAR_NAME = core.kTempVarName()
 GRAD_VAR_SUFFIX = core.kGradVarSuffix()
 ZERO_VAR_SUFFIX = core.kZeroVarSuffix()
+CONTROL_DEP_VAR_PREFIX = core.kControlDepVarName()
+
+
+def generate_control_dev_var_name():
+    import random
+    return CONTROL_DEP_VAR_PREFIX + "@" + str(random.random())
 
 
 def grad_var_name(var_name):

python/paddle/fluid/transpiler/distribute_transpiler.py

@@ -211,8 +211,10 @@ class DistributeTranspiler(object):
         ps_dispatcher = self.config.split_method(self.pserver_endpoints)
         self.has_distributed_lookup_table = self._has_distributed_lookup_table()
         self.param_name_to_grad_name = dict()
+        self.grad_name_to_param_name = dict()
         for param_var, grad_var in self.params_grads:
             self.param_name_to_grad_name[param_var.name] = grad_var.name
+            self.grad_name_to_param_name[grad_var.name] = param_var.name
 
         # step 1: split and create vars, then put splited vars in dicts for later use.
         self._init_splited_vars()
@@ -254,8 +256,10 @@ class DistributeTranspiler(object):
                 AssertionError("Can not insert the send op by original "
                                "variable name :", splited_grad_varname)
 
-            dummy_output = program.global_block().create_var()
+            dummy_output = program.global_block().create_var(
+                name=framework.generate_control_dev_var_name())
             grad_name_to_send_dummy_out[grad_varname] = dummy_output
+
             program.global_block()._insert_op(
                 index=index + 1,
                 type="send",
@@ -264,6 +268,8 @@ class DistributeTranspiler(object):
                 attrs={
                     "epmap": eplist,
                     RPC_OP_ROLE_ATTR_NAME: RPC_OP_ROLE_ATTR_VALUE,
+                    OP_ROLE_VAR_ATTR_NAME:
+                    [self.grad_name_to_param_name[grad_varname], grad_varname],
                     "sync_mode": not self.sync_mode,
                 })
             for _, var in enumerate(splited_vars):
@@ -305,6 +311,10 @@ class DistributeTranspiler(object):
                 attrs={
                     "epmap": eps,
                     RPC_OP_ROLE_ATTR_NAME: RPC_OP_ROLE_ATTR_VALUE,
+                    OP_ROLE_VAR_ATTR_NAME: [
+                        param_varname,
+                        self.param_name_to_grad_name[param_varname]
+                    ],
                     "sync_mode": not self.sync_mode
                 })
 
@@ -934,7 +944,11 @@ class DistributeTranspiler(object):
                     attrs={
                         "sync_mode": True,
                         "epmap": pserver_endpoints,
-                        RPC_OP_ROLE_ATTR_NAME: RPC_OP_ROLE_ATTR_VALUE
+                        RPC_OP_ROLE_ATTR_NAME: RPC_OP_ROLE_ATTR_VALUE,
+                        OP_ROLE_VAR_ATTR_NAME: [
+                            self.grad_name_to_param_name[table_grad_name],
+                            table_grad_name
+                        ]
                     })
                 break