提交 da7b9a5e 编写于 作者: D dangqingqing

Remove the warning in hl_batch_norm_forward_inference function.

上级 3e4c671c
...@@ -1023,14 +1023,6 @@ void hl_batch_norm_forward_inference(hl_tensor_descriptor inputDesc, ...@@ -1023,14 +1023,6 @@ void hl_batch_norm_forward_inference(hl_tensor_descriptor inputDesc,
real beta = 1.0f; real beta = 1.0f;
cudnnBatchNormMode_t mode = CUDNN_BATCHNORM_SPATIAL; 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( CHECK_CUDNN(
dynload::cudnnBatchNormalizationForwardInference(t_resource.cudnn_handle, dynload::cudnnBatchNormalizationForwardInference(t_resource.cudnn_handle,
mode, mode,
......
Markdown is supported
0% .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑!
想要评论请 注册