[GitHub] eric-haibin-lin commented on a change in pull request #14240: use safe accumulation for norm

[GitBox <gi...@apache.org>]

eric-haibin-lin commented on a change in pull request #14240: use safe accumulation for norm
URL: https://github.com/apache/incubator-mxnet/pull/14240#discussion_r261009149
 
 

 ##########
 File path: src/operator/tensor/broadcast_reduce-inl.cuh
 ##########
 @@ -610,14 +609,22 @@ void ReduceImpl(cudaStream_t stream, const TBlob& small, const TBlob& lhs, const
 
 #undef KERNEL_UNROLL_SWITCH
 
-template<typename Reducer, int ndim, typename DType, typename OP>
+template<typename Reducer, int ndim, typename DType, typename OP, bool safe_acc = false>
 void Reduce(Stream<gpu> *s, const TBlob& small, const OpReqType req,
             const Tensor<gpu, 1, char>& workspace, const TBlob& big) {
   if (req == kNullOp) return;
   cudaStream_t stream = Stream<gpu>::GetStream(s);
   ReduceImplConfig<ndim> config =
     ConfigureReduceImpl<ndim, DType>(small.shape_, big.shape_, NULL, NULL);
-  ReduceImpl<Reducer, ndim, DType, OP>(stream, small, req, big, workspace, config);
+  if (safe_acc) {
+    MXNET_REAL_ACC_TYPE_SWITCH(mshadow::DataType<DType>::kFlag, DataType, AType, {
 
 Review comment:
   Need to support acc type for int, etc

----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on GitHub and use the
URL above to go to the specific comment.
 
For queries about this service, please contact Infrastructure at:
users@infra.apache.org


With regards,
Apache Git Services