From 92bcdcf5b0afe400d6a866da000a1df5d8519d68 Mon Sep 17 00:00:00 2001
From: Wenkai Du <43822138+wenkaidu@users.noreply.github.com>
Date: Thu, 20 May 2021 08:58:45 -0700
Subject: [PATCH] Correction on max number of groups (#373)

[ROCm/rccl commit: 50da1b48afc61f5da444bd487bd289d2966dca17]
---
 projects/rccl/src/include/devcomm.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/projects/rccl/src/include/devcomm.h b/projects/rccl/src/include/devcomm.h
index 4c7c9980bc..3de95ef0d5 100644
--- a/projects/rccl/src/include/devcomm.h
+++ b/projects/rccl/src/include/devcomm.h
@@ -176,7 +176,7 @@ struct ncclDevComm;
 #pragma pack(push)  /* push current alignment to stack */
 #pragma pack(4)     /* set alignment to 4 bytes boundary */
 #define NCCL_MAX_WORK_ELEMENTS 1
-#define NCCL_MAX_GROUPS (NCCL_MAX_WORK_ELEMENTS*2)
+#define NCCL_MAX_GROUPS (NCCL_MAX_NTHREADS/WARP_SIZE)
 
 /* ncclWork is to be a power of two, currently 8x64 bytes, */
 /* to make sure reads to host from the CUDA kernel are aligned. */