Tune inception (#1014)

* let NormalMdUpdtTask use independent stream

* let kMix task share the same stream

* remove UseIndependentStream

* refine

oneflow/core/actor/actor.cpp

@@ -121,13 +121,8 @@ void Actor::InitDeviceCtx(const ThreadCtx& thread_ctx) {
     }
     case DeviceType::kGPU: {
       CudaStreamHandle* cuda_handle = nullptr;
-      if (GetLocalWorkStreamId() == 0) {
-        cuda_handle = thread_ctx.g_cuda_stream.get();
-      } else {
-        CHECK(Global<IDMgr>::Get()->IsIndependentLocalWorkStreamId(GetLocalWorkStreamId()));
-        cuda_handle_.reset(new CudaStreamHandle(thread_ctx.cb_event_chan));
-        cuda_handle = cuda_handle_.get();
-      }
+      CHECK_EQ(GetLocalWorkStreamId(), 0);
+      cuda_handle = thread_ctx.g_cuda_stream.get();
       device_ctx_.reset(new CudaDeviceCtx(thread_ctx.buf_ptr, thread_ctx.buf_size, cuda_handle));
       break;
     }

oneflow/core/graph/compute_task_node.h

@@ -16,7 +16,6 @@ class CompTaskNode : public TaskNode {
 
   virtual CudaWorkType GetCudaWorkType() const { return CudaWorkType::kCompute; }
   virtual void ToProto(TaskProto*) override;
-  bool UseIndependentWorkStream() const override { return GetCudaWorkType() == CudaWorkType::kMix; }
 
   // parallel_ctx_
   int64_t parallel_id() const { return parallel_ctx_.parallel_id(); }

oneflow/core/graph/normal_model_update_compute_task_node.h

@@ -19,6 +19,7 @@ class NormalMdUpdtCompTaskNode final : public CompTaskNode {
 
   void set_random_seed(uint32_t val) { random_seed_ = val; }
   TaskType GetTaskType() const override { return TaskType::kNormalMdUpdt; }
+  CudaWorkType GetCudaWorkType() const override { return CudaWorkType::kMix; }
   void ToProto(TaskProto*) override;
 
  private:

oneflow/core/graph/task_node.cpp

@@ -266,17 +266,7 @@ void TaskNode::FixRegisterNumRange() {
 int64_t TaskNode::AllocateLocalWorkStreamId() {
   CHECK_NE(machine_id_, -1);
   CHECK_NE(thrd_id_, -1);
-  if (UseIndependentWorkStream()) {
-    if (device_type() == DeviceType::kCPU) {
-      return 0;
-    } else if (device_type() == DeviceType::kGPU) {
-      return Global<IDMgr>::Get()->AllocateLocalWorkStreamId(machine_id_, thrd_id_);
-    } else {
-      UNIMPLEMENTED();
-    }
-  } else {
-    return 0;
-  }
+  return 0;
 }
 
 void TaskNode::UpdateTaskId() {

oneflow/core/graph/task_node.h

@@ -49,7 +49,6 @@ class TaskNode : public Node<TaskNode, TaskEdge> {
   int64_t LocalWorkStreamId() const;
   int64_t GlobalWorkStreamId() const;
   int64_t GpuPhyId() const { return Global<IDMgr>::Get()->GetGpuPhyIdFromThrdId(thrd_id_); }
-  virtual bool UseIndependentWorkStream() const { return false; }
 
   // Setters
   void set_machine_id(int64_t val);

oneflow/core/job/id_manager.h

@@ -47,11 +47,9 @@ class IDMgr final {
   //   0: the actor thread
   // for gpu:
   //   0: the global cuda stream
-  //   other: start from 100
   int64_t AllocateLocalWorkStreamId(int64_t machine_id, int64_t thrd_id);
   int64_t LocalWorkStreamId4TaskId(int64_t task_id) const;
   int64_t LocalWorkStreamId4ActorId(int64_t actor_id) const;
-  bool IsIndependentLocalWorkStreamId(int64_t local_wsid) const { return local_wsid >= 100; }
   // global_work_stream_id
   // sign | machine_id | thrd_id | local_work_stream_id | 0
   //  1   |     10     |   11    |          21          | 21