fix ci bug: cudnn handler in multi card

test=develop

paddle/fluid/platform/device_context.cc

@@ -212,6 +212,7 @@ class EigenCudaStreamDevice : public Eigen::StreamInterface {
 
 CudnnHolder::CudnnHolder(const cudaStream_t* stream, const CUDAPlace& place)
     : workspace_(nullptr), stream_(stream), place_(place) {
+  PADDLE_ENFORCE(cudaSetDevice(place_.device));
   PADDLE_ENFORCE(dynload::cudnnCreate(&cudnn_handle_));
   PADDLE_ENFORCE(dynload::cudnnSetStream(cudnn_handle_, *stream_));
 }
@@ -233,8 +234,6 @@ void CudnnHolder::ReallocateWorkspace(size_t required_workspace_len) {
                                      paddle::memory::Allocator::kScratchpad);
 }
 
-std::once_flag CUDADeviceContext::init_cudnn_;
-
 CUDADeviceContext::CUDADeviceContext(CUDAPlace place)
     : place_(place), cudnn_holder_(nullptr) {
   CUDADeviceGuard guard(place_.device);

paddle/fluid/platform/device_context.h

@@ -292,7 +292,7 @@ class CUDADeviceContext : public DeviceContext {
  private:
   CUDAPlace place_;
 
-  static std::once_flag init_cudnn_;
+  mutable std::once_flag init_cudnn_;
 
   std::unique_ptr<Eigen::GpuDevice> eigen_device_;
   std::unique_ptr<EigenCudaStreamDevice> eigen_stream_;