AI/rocm-systems
2
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Only use split tree algorithm to reduce kernel code size.

Browse Source
This commit is contained in:
Wenkai Du
2022-09-20 18:35:26 +00:00
committed by Wen-Heng (Jack) Chung
parent 49c811ecf9
commit 02929cffb6
1 changed files with 1 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/collectives/device/all_reduce.h
+1 -2
View File
@@ -691,8 +691,7 @@ struct RunWorkElement<ncclFuncAllReduce, T, RedOp, NCCL_ALGO_RING, NCCL_PROTO_LL
template<typename T, typename RedOp>
struct RunWorkElement<ncclFuncAllReduce, T, RedOp, NCCL_ALGO_TREE, NCCL_PROTO_LL> {
__device__ __forceinline__ void run(ncclWorkElem *args) {
if (args->pad_0 == 0) runTreeUpDown<T, RedOp, ProtoLL>(args);
else runTreeSplit<T, RedOp, ProtoLL>(args);
runTreeSplit<T, RedOp, ProtoLL>(args);
}
};