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

Increase chunk steps of broadcast and reduce (#528)

Browse Source 
[ROCm/rccl commit: 15b572751e]
This commit is contained in:
Wenkai Du
2022-04-07 13:34:04 -07:00
committed by GitHub
parent 3830310156
commit 5ccdd9f5e1
1 changed files with 3 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files
projects/rccl/src/include/collectives.h
+3 -3
View File
@@ -123,11 +123,11 @@ extern __device__ void NCCL_ONERANK_REDUCE_NAME(PreMulSum, double)(struct ncclWo
#define REDUCESCATTER_SLICESTEPS 2
#define REDUCESCATTER_CHUNKSTEPS 4
#define BROADCAST_SLICESTEPS 1
#define BROADCAST_CHUNKSTEPS 2
#define BROADCAST_CHUNKSTEPS 4
#define REDUCE_SLICESTEPS 1
#define REDUCE_CHUNKSTEPS 2
#define REDUCE_CHUNKSTEPS 4
#define SENDRECV_SLICEFACTOR 1
#define NCCL_MAX_SLICE_PER_CHUNK 2 // max value for CHUNKSTEPS/SLICESTEPS, must accord with above
#define NCCL_MAX_SLICE_PER_CHUNK 4 // max value for CHUNKSTEPS/SLICESTEPS, must accord with above
#define ALLTOALL_PIVOT_SLICESTEPS 2
#define ALLTOALL_PIVOT_CHUNKSTEPS 4