Remove global variable in CHECK_CUDA and CHECK_CUDNN and fix compiler bug.

ISSUE=4586769

git-svn-id: https://svn.baidu.com/idl/trunk/paddle@1409 1ad973e4-5ce8-4261-8a94-b56d1f490c56

paddle/cuda/src/hl_cuda_cudnn.cc

@@ -175,15 +175,16 @@ CUDNN_DNN_ROUTINE_EACH_R5(DYNAMIC_LOAD_CUDNN_WRAP)
 } /* namespace dynload */
 
 /**
- * Check build-in cudnn function using glog and it also
+ * Check build-in cudnn function using glog and it **does not**
  * support << operator for more details error info.
  */
-cudnnStatus_t g_cudnnStat;
+#define CHECK_CUDNN(cudnnFunc)                               \
-#define CHECK_CUDNN(cudnn_func)                                \
+  do {                                                       \
-  g_cudnnStat = cudnn_func;                                    \
+    cudnnStatus_t cudnnStat = cudnnFunc;                     \
-  CHECK_EQ(CUDNN_STATUS_SUCCESS, g_cudnnStat)                  \
+    CHECK_EQ(CUDNN_STATUS_SUCCESS, cudnnStat)                \
-      << "Cudnn Error: "                                       \
+        << "Cudnn Error: "                                   \
-      << dynload::cudnnGetErrorString(g_cudnnStat) << ". "     \
+        << dynload::cudnnGetErrorString(cudnnStat);          \
+  } while (0)
 
 bool g_is_libcudnn_init = false;
 int g_cudnn_lib_version = 0;

paddle/cuda/src/hl_cuda_device.cc

@@ -178,14 +178,15 @@ int g_cuda_lib_version = 0;
 #define  HPPL_GPU_MEMORY_SIZE                (256*4)
 
 /**
- * Check build-in cuda function using glog and it also
+ * Check build-in cuda function using glog and it **does not**
  * support << operator for more details error info.
  */
-cudaError_t cudaStat;
+#define CHECK_CUDA(cudaFunc)                               \
-#define CHECK_CUDA(cuda_func)                                 \
+  do {                                                     \
-  cudaStat = cuda_func;                                     \
+    cudaError_t cudaStat = cudaFunc;                       \
-  CHECK_EQ(cudaSuccess, cudaStat) << "Cuda Error: "         \
+    CHECK_EQ(cudaSuccess, cudaStat) << "Cuda Error: "      \
-      << dynload::cudaGetErrorString(cudaStat) << ". "      \
+        << dynload::cudaGetErrorString(cudaStat);          \
+  } while (0)
 
 /**
  * @brief   thread resource.

paddle/trainer/RemoteParameterUpdater.h

@@ -29,7 +29,7 @@ namespace paddle {
  * parameters servers.
  */
 
-// TODO:(yanfei)
+// TODO(yanfei):
 // I think that the biggest feature of rdma is packet lossless control
 // feature instead of high bandwiths, zero copy and gpu-direct rdma in
 // theroy.
@@ -164,7 +164,7 @@ protected:
   static const std::string kElasticAverage;
 };
 
-// TODO:(yanfei)
+// TODO(yanfei):
 // do parameters level synchronization Optimization at pserver end with
 // ConcurrentRemoteParameterUpdater to get more parallelization, at last
 // to really hide pserver latency in backward computation.
@@ -241,7 +241,7 @@ private:
   bool oneBatchFinished_;
 };
 
-// TODO:(yanfei)
+// TODO(yanfei):
 // merge sparse updater with dense updater, and could help to reduce
 // the synchronization between sparse and dense udpater. it could also
 // reduce the threads for managing all connections.

paddle/trainer/Trainer.cpp

@@ -188,7 +188,7 @@ void Trainer::init(const std::shared_ptr<TrainerConfigHelper> &config,
 
 
   bool gpuData = FLAGS_use_gpu && (!FLAGS_parallel_nn) &&
-                 (!IGradientMachineMode::dataMustInCpu(mode_, 
+                 (!IGradientMachineMode::dataMustInCpu(mode_,
                                                        FLAGS_trainer_count));
 
   dataProvider_ = dataProvider;