2
0

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

* Update LL128 elems per thread

* Precompute ix[g] in LL128 prim

* Make Threadthreshold part of tuning models

* Ignore channel tuning when channels are env controlled

* Tune LL128 max limit for AG

* Tune LL128 max limit for RS

* Retune AR LL128 limits due to changes

* Update CHANGELOG.md

---------

Co-authored-by: Jeffrey Novotny <jnovotny@amd.com>

[ROCm/rccl commit: 00c1eb098c]
Este cometimento está contido em:
Mustafa Abduljabbar
2025-05-14 14:35:54 -05:00
cometido por GitHub
ascendente 3731cae1b7
cometimento 951ed9cde1
7 ficheiros modificados com 52 adições e 15 eliminações
+2 -2
Ver ficheiro
@@ -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 16
#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);