From abdc365a058d8c368763f829babda4bb2abf07c3 Mon Sep 17 00:00:00 2001
From: Wenkai Du <wenkai.du@amd.com>
Date: Tue, 20 Sep 2022 18:35:26 +0000
Subject: [PATCH] Only use split tree algorithm to reduce kernel code size.

[ROCm/rccl commit: 02929cffb6fb594ba13da461c190edced82c06f7]
---
 projects/rccl/src/collectives/device/all_reduce.h | 3 +--
 1 file changed, 1 insertion(+), 2 deletions(-)

diff --git a/projects/rccl/src/collectives/device/all_reduce.h b/projects/rccl/src/collectives/device/all_reduce.h
index f71953aa26..465fe32f54 100644
--- a/projects/rccl/src/collectives/device/all_reduce.h
+++ b/projects/rccl/src/collectives/device/all_reduce.h
@@ -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);
   }
 };