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

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

Browse Source 
* 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]
This commit is contained in:
Pedram Alizadeh
2025-06-04 13:22:11 -04:00
committed by GitHub
parent a50ff2c3d3
commit 1ace5d05ed
7 changed files with 50 additions and 17 deletions
Download Patch File Download Diff File Expand all files Collapse all files
projects/rccl/src/include/rccl_common.h
+3 -3
View File
@@ -33,10 +33,11 @@ typedef enum RcclTunableColls {
} rcclTunableIndex_t;
#define RCCL_LL_LIMITS_UNDEFINED 0
#define RCCL_PROTOCOL_ENTRY_SIZE 3
#define RCCL_PROTOCOL_ENTRY_SIZE 4
#define RCCL_PROTOCOL_MIN_IDX 0
#define RCCL_PROTOCOL_MAX_IDX 1
#define RCCL_PROTOCOL_FACTOR_IDX 2
#define RCCL_PROTOCOL_THREAD_THRESHOLD_IDX 3
#ifdef RCCL_EXPOSE_STATIC
#define RCCL_STATIC_EXPOSE_CHECK()
@@ -71,8 +72,7 @@ inline size_t rcclGetSizePerRank(ncclFunc_t const& func, size_t const& nBytes, i
return (func == ncclFuncReduceScatter || func == ncclFuncAllGather) ? nBytes / nRanks : nBytes;
}
void rcclUpdateCollectiveProtocol(struct ncclComm* comm, size_t const& nBytes, struct ncclTaskColl* info);
void rcclUpdateThreadThreshold(struct ncclComm* comm, size_t const& nBytes, struct ncclTaskColl* info, int& threadThreshold);
ncclResult_t rcclGetAlgoInfo(struct ncclComm* comm, ncclFunc_t coll, uint64_t count, ncclDataType_t dataType,
int collNetSupport, int nvlsSupport, int numPipeOps,
int* algo, int* protocol, int* maxChannels);