Tune collectives for 2.11.4 (#486)

[ROCm/rccl commit: 565fbeb5e9]
This commit is contained in:
Wenkai Du
2022-01-10 08:25:47 -08:00
committed by GitHub
parent a6a9ba1b78
commit 3f16835565
+7 -7
View File
@@ -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