Merge pull request #3185 from kuke/check_cudnn_dev

Add the check of cudnn version in cudnnBatchNorm

paddle/cuda/src/hl_cuda_cudnn.cc

@@ -1022,6 +1022,15 @@ void hl_batch_norm_forward_inference(hl_tensor_descriptor inputDesc,
   real alpha = 1.0f;
   real beta = 1.0f;
   cudnnBatchNormMode_t mode = CUDNN_BATCHNORM_SPATIAL;
+
+  int batch_size = ((cudnn_tensor_descriptor)inputDesc)->batch_size;
+  if (batch_size > 1024 && g_cudnn_lib_version < 6000) {
+    LOG(INFO) << " To process current batch data with size " << batch_size
+              << " (>1024), cudnnBatchNorm requires cuDNN version >= 6000."
+              << " If there is an error complaining CUDNN_STATUS_NOT_SUPPORTED,"
+              << " just recompile PaddlePaddle with cuDNN >= 6000, replacing"
+              << " current version " << g_cudnn_lib_version;
+  }
   CHECK_CUDNN(
       dynload::cudnnBatchNormalizationForwardInference(t_resource.cudnn_handle,
                                                        mode,