Avoid init_nccl for every steps.

paddle/fluid/operators/nccl/nccl_gpu_common.cc

@@ -16,5 +16,44 @@ limitations under the License. */
 #include "paddle/fluid/platform/gpu_info.h"
 
 namespace paddle {
-namespace platform {}  // namespace platform
+namespace platform {
+namespace {
+// TODO(panyx0718): Where to destroy them.
+std::unique_ptr<std::vector<ncclComm_t>> global_comms;
+std::unique_ptr<std::unordered_map<int, int>> comm_id_map;
+bool inited = false;
+size_t last_num_gpus = -1;
+}
+
+int Communicator::GetCommId(int device_id) const {
+  return comm_id_map->at(device_id);
+}
+
+void Communicator::InitAll(const std::vector<int>& gpus) {
+  if (inited && last_num_gpus == gpus.size()) {
+    return;
+  }
+  last_num_gpus = gpus.size();
+  if (global_comms) {
+    for (size_t i = 0; i < global_comms->size(); ++i) {
+      // FIXME(dzh) : PADDLE_ENFORCE return void
+      dynload::ncclCommDestroy((*global_comms)[i]);
+    }
+  }
+  global_comms.reset(new std::vector<ncclComm_t>());
+  comm_id_map.reset(new std::unordered_map<int, int>());
+  global_comms->resize(gpus.size());
+  for (size_t i = 0; i < gpus.size(); ++i) {
+    (*comm_id_map)[gpus[i]] = i;
+  }
+  PADDLE_ENFORCE(
+      dynload::ncclCommInitAll(global_comms->data(), gpus.size(), gpus.data()));
+  inited = true;
+}
+
+const std::vector<ncclComm_t>& Communicator::comms() const {
+  return *global_comms;
+}
+
+}  // namespace platform
 }  // namespace paddle

paddle/fluid/operators/nccl/nccl_gpu_common.h

@@ -29,39 +29,16 @@ limitations under the License. */
 
 namespace paddle {
 namespace platform {
-
 constexpr int kInvalidGPUId = -1;
 
 struct Communicator {
-  std::vector<ncclComm_t> comms_;
-  std::unordered_map<int, int> comm_id_map_;
-  bool inited_;
-
   Communicator() {}
 
-  int GetCommId(int device_id) const { return comm_id_map_.at(device_id); }
+  int GetCommId(int device_id) const;
-
-  void InitAll(const std::vector<int>& gpus) {
-    comms_.resize(gpus.size());
-    inited_ = false;
-    for (size_t i = 0; i < gpus.size(); ++i) {
-      comm_id_map_[gpus[i]] = i;
-    }
-    PADDLE_ENFORCE(
-        dynload::ncclCommInitAll(comms_.data(), gpus.size(), gpus.data()));
-    inited_ = true;
-  }
 
-  ~Communicator() {
+  void InitAll(const std::vector<int>& gpus);
-    if (inited_) {
-      for (size_t i = 0; i < comms_.size(); ++i) {
-        // FIXME(dzh) : PADDLE_ENFORCE return void
-        dynload::ncclCommDestroy(comms_[i]);
-      }
-    }
-  }
 
-  DISABLE_COPY_AND_ASSIGN(Communicator);
+  const std::vector<ncclComm_t>& comms() const;
 };
 
 }  // namespace platform

paddle/fluid/operators/nccl_op.cu.cc

@@ -78,7 +78,7 @@ class NCCLAllReduceKernel : public framework::OpKernel<T> {
       PADDLE_ENFORCE(platform::dynload::ncclAllReduce(
           ins[i]->data<T>(), outs[i]->mutable_data<T>(ctx.GetPlace()),
           outs[i]->numel(), NCCLTypeWrapper<T>::type, reduction_op_,
-          comm->comms_[idx], stream));
+          comm->comms().at(idx), stream));
       PADDLE_ENFORCE(cudaStreamSynchronize(stream));
 
       VLOG(1) << "gpu : "
@@ -127,7 +127,7 @@ class NCCLReduceKernel : public framework::OpKernel<T> {
     std::hash<std::string> hasher;
     for (size_t i = 0; i < ins.size(); ++i) {
       if (root == platform::kInvalidGPUId) {
-        root = hasher(ins_names[i]) % comm->comms_.size();
+        root = hasher(ins_names[i]) % comm->comms().size();
       }
       T* recvbuffer = nullptr;
       if (root == gpu_id) {
@@ -139,7 +139,7 @@ class NCCLReduceKernel : public framework::OpKernel<T> {
 
       PADDLE_ENFORCE(platform::dynload::ncclReduce(
           ins[i]->data<T>(), recvbuffer, ins[i]->numel(),
-          NCCLTypeWrapper<T>::type, reduction_op_, root, comm->comms_[idx],
+          NCCLTypeWrapper<T>::type, reduction_op_, root, comm->comms().at(idx),
           stream));
       PADDLE_ENFORCE(cudaStreamSynchronize(stream));
 
@@ -176,7 +176,7 @@ class NCCLBcastKernel : public framework::OpKernel<T> {
         VLOG(1) << " before ncclBcast";
         PADDLE_ENFORCE(platform::dynload::ncclBcast(
             (void*)ins[i]->data<T>(), ins[i]->numel(), NCCLTypeWrapper<T>::type,
-            root, comm->comms_[idx], stream));
+            root, comm->comms().at(idx), stream));
         VLOG(1) << " after ncclBcast";
         PADDLE_ENFORCE(cudaStreamSynchronize(stream));
 
@@ -190,7 +190,7 @@ class NCCLBcastKernel : public framework::OpKernel<T> {
 
         PADDLE_ENFORCE(platform::dynload::ncclBcast(
             outs[i]->mutable_data<T>(ctx.GetPlace()), outs[i]->numel(),
-            NCCLTypeWrapper<T>::type, root, comm->comms_[idx], stream));
+            NCCLTypeWrapper<T>::type, root, comm->comms().at(idx), stream));
         PADDLE_ENFORCE(cudaStreamSynchronize(stream));
 
         VLOG(1) << "gpu : " << gpu_id << " finished Bcast. recv "