Support control flow in DataParallel

paddle/fluid/framework/distributed_strategy.proto

@@ -151,6 +151,7 @@ message DistributedStrategy {
   optional bool fp16_allreduce = 25 [ default = false ];
   optional bool sharding = 26 [ default = false ];
   optional float last_comm_group_size_MB = 27 [ default = 1 ];
+  optional bool find_unused_parameters = 28 [ default = true ];
 
   optional RecomputeConfig recompute_configs = 101;
   optional AMPConfig amp_configs = 102;

paddle/fluid/imperative/bkcl_context.cc

@@ -153,6 +153,14 @@ void BKCLParallelContext::WaitCompute(int ring_id) {
   compute_dev_ctx->Wait();
 }
 
+void BKCLParallelContext::SynchronizeCompute() {
+  // TODO(wangxi16): [Performance optimize] Maybe need to put Wait and
+  // bkcl_allreduce to comm thread, for bkcl_allreduce is blocking now.
+  auto compute_dev_ctx = static_cast<platform::XPUDeviceContext *>(
+      platform::DeviceContextPool::Instance().Get(place_));
+  compute_dev_ctx->Wait();
+}
+
 void BKCLParallelContext::WaitComm(int ring_id) {
   PADDLE_ENFORCE_GE(ring_id, 0,
                     platform::errors::OutOfRange(

paddle/fluid/imperative/bkcl_context.h

@@ -45,6 +45,8 @@ class BKCLParallelContext : public ParallelContext {
   void WaitCompute(int ring_id) override;
 
   void WaitComm(int ring_id) override;
+
+  void SynchronizeCompute() override;
 };
 
 }  //  namespace imperative

paddle/fluid/imperative/nccl_context.cc

@@ -149,6 +149,12 @@ void NCCLParallelContext::WaitComm(int ring_id) {
 #endif
 }
 
+void NCCLParallelContext::SynchronizeCompute() {
+  auto *compute_dev_ctx = static_cast<platform::CUDADeviceContext *>(
+      platform::DeviceContextPool::Instance().Get(place_));
+  compute_dev_ctx->Wait();
+}
+
 #endif
 
 }  //  namespace imperative

paddle/fluid/imperative/nccl_context.h

@@ -63,6 +63,8 @@ class NCCLParallelContext : public ParallelContext {
 
   void WaitComm(int ring_id) override;
 
+  void SynchronizeCompute() override;
+
  private:
   // used for comm wait compute, compute_stream-->event-->comm_stream[ring_id]
   std::vector<std::shared_ptr<platform::CudaEventObject>> compute_events_;

paddle/fluid/imperative/parallel_context.h

@@ -64,6 +64,9 @@ class ParallelContext {
   // if CPU, should do nothing.
   virtual void WaitComm(int ring_id) = 0;
 
+  // synchorize compute stream
+  virtual void SynchronizeCompute() = 0;
+
   inline int GetNRings() const { return strategy_.nrings_; }
 
   inline int64_t GetNRanks() const { return strategy_.nranks_; }