From 325ff1dc1129fa2bf7d0a885ded366c5c0a57fd5 Mon Sep 17 00:00:00 2001 From: Nusrat Islam Date: Thu, 25 May 2023 15:42:18 -0500 Subject: [PATCH] device: change unroll factor The default value of unroll factor is 2. Changing the unroll factor to 4 provides better performance for most of the collectives. [ROCm/rccl commit: 4d1cfb17c8d591356bf4f73f2013952f28ad94c1] --- projects/rccl/src/collectives/device/common.h | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/projects/rccl/src/collectives/device/common.h b/projects/rccl/src/collectives/device/common.h index 0d9cd3b734..a7ebcad5b1 100644 --- a/projects/rccl/src/collectives/device/common.h +++ b/projects/rccl/src/collectives/device/common.h @@ -11,7 +11,7 @@ #include "collectives.h" #include "devcomm.h" -#define COLL_UNROLL 2 +#define COLL_UNROLL 4 #define NCCL_MAX_DEV_ARITY (NCCL_MAX_TREE_ARITY-1) // Using balanced tree instead of split tree #define __syncwarp()