Reapplying PR #1641 [AG and RS channel tuning] Add thread work threshold to tuning models and precompute reg index in LL128 (#1713)

* Reapply "[AG and RS channel tuning] Add thread work threshold to tuning models and precompute reg index in LL128 (#1641)"

This reverts commit 943ad6f7820739385a0b54e81f823d0df1dbf71c.

* Decreasing NCCL_LL128_SHMEM_ELEMS_PER_THREAD from 16 to 8

[ROCm/rccl commit: 3f7c08648f]
Этот коммит содержится в:
Pedram Alizadeh
2025-06-04 13:22:11 -04:00
коммит произвёл GitHub
родитель a50ff2c3d3
Коммит 1ace5d05ed
7 изменённых файлов: 50 добавлений и 17 удалений
+2 -2
Просмотреть файл
@@ -95,7 +95,7 @@ static_assert(NCCL_LL_CLEAN_MASK % NCCL_STEPS == 0, "Invalid NCCL_LL_CLEAN_MASK
#define NCCL_LL128_MAX_NTHREADS 256
#define NCCL_LL128_ELEMS_PER_THREAD 28
#define NCCL_LL128_SHMEM_ELEMS_PER_THREAD 4
#define NCCL_LL128_SHMEM_ELEMS_PER_THREAD 8
#define NCCL_LL128_SHMEM_SIZE (NCCL_LL128_SHMEM_ELEMS_PER_THREAD*NCCL_LL128_MAX_NTHREADS)
#define NCCL_P2P_WRITE 0x01
@@ -698,7 +698,7 @@ inline int ncclDevFuncId(int coll, int devRedOp, int type, int algo, int proto)
// RING / <all_protos> / <all_redops> / <all_types>
if (coll == ncclFuncReduce) {
row += ((proto * ncclNumDevRedOps + devRedOp) * ncclNumTypes + type) - NCCL_NUM_FLOATS * proto;
row += ((proto * ncclNumDevRedOps + devRedOp) * ncclNumTypes + type) - NCCL_NUM_FLOATS * proto;
break;
}
row += NCCL_NUM_PROTOCOLS * (ncclNumDevRedOps * ncclNumTypes - NCCL_NUM_FLOATS);