From 3f16835565b6178a90ae3c4fe188715b4a16fd86 Mon Sep 17 00:00:00 2001
From: Wenkai Du <43822138+wenkaidu@users.noreply.github.com>
Date: Mon, 10 Jan 2022 08:25:47 -0800
Subject: [PATCH] Tune collectives for 2.11.4 (#486)

[ROCm/rccl commit: 565fbeb5e9b6a9c7668498193648ac12ede6f7db]
---
 projects/rccl/src/include/collectives.h | 14 +++++++-------
 1 file changed, 7 insertions(+), 7 deletions(-)

diff --git a/projects/rccl/src/include/collectives.h b/projects/rccl/src/include/collectives.h
index 1d0c56cf54..b65e81b3ea 100644
--- a/projects/rccl/src/include/collectives.h
+++ b/projects/rccl/src/include/collectives.h
@@ -114,16 +114,16 @@ extern __device__ void NCCL_ONERANK_REDUCE_NAME(PreMulSum, double)(struct ncclWo
 //#define ALLGATHER_CHUNKSTEPS (NCCL_STEPS/2)
 //#define REDUCESCATTER_SLICESTEPS (NCCL_STEPS/4)
 //#define REDUCESCATTER_CHUNKSTEPS (NCCL_STEPS/2)
-#define ALLREDUCE_SLICESTEPS 4
-#define ALLREDUCE_CHUNKSTEPS 4
-#define ALLGATHER_SLICESTEPS 4
-#define ALLGATHER_CHUNKSTEPS 4
-#define REDUCESCATTER_SLICESTEPS 4
+#define ALLREDUCE_SLICESTEPS 1
+#define ALLREDUCE_CHUNKSTEPS 1
+#define ALLGATHER_SLICESTEPS 1
+#define ALLGATHER_CHUNKSTEPS 1
+#define REDUCESCATTER_SLICESTEPS 2
 #define REDUCESCATTER_CHUNKSTEPS 4
 #define BROADCAST_SLICESTEPS 1
-#define BROADCAST_CHUNKSTEPS 1
+#define BROADCAST_CHUNKSTEPS 2
 #define REDUCE_SLICESTEPS 1
-#define REDUCE_CHUNKSTEPS 1
+#define REDUCE_CHUNKSTEPS 2
 #define SENDRECV_SLICEFACTOR 1
 #define NCCL_MAX_SLICE_PER_CHUNK 2  // max value for CHUNKSTEPS/SLICESTEPS, must accord with above