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

[Device] Adjust threadblock size for gfx950 to increase LL64/Simple performance for AR, RS and AG (#1978)

Browse Source 
* Add initial commit to increase tb size to 512
* Fix LL perf issue when subset of NCCL_MAX_NTHREADS is used
Adding a constant to barrier_generic logic from using fallback logic when nthreads < NCCL_MAX_NTHREADS and nthreads == blockDim.X
* Adjust nthreads for LL
* Opt threads for reduce_scatter upper small range
* Add macro for single node
* Restrict MSCCL to 256 threads to prevent mem access fault
* Support pre-MI350 compatibility
* Partially refactor threadblock size override
* Use const macros instead of numerals
* opt out of unused function

[ROCm/rccl commit: 12f51ba8bf]
This commit is contained in:
Mustafa Abduljabbar
2025-10-30 00:24:32 -04:00
committed by GitHub
parent 4c7afea115
commit eb0b1387b7
13 changed files with 120 additions and 40 deletions
Download Patch File Download Diff File Expand all files Collapse all files
projects/rccl/src/misc/msccl/msccl_setup.cc
+1 -1
View File
@@ -420,7 +420,7 @@ ncclResult_t mscclSetupKernel(const void* sendBuff, void* recvBuff, size_t count
uint32_t numBlocks = (uint32_t)hostAlgo->nBlocks;
dim3 grid = {numBlocks, 1, 1};
dim3 block = {NCCL_MAX_NTHREADS, 1, 1};
dim3 block = {MSCCL_MAX_NTHREADS, 1, 1};
ncclDevRedOpFull opFull = {};
NCCLCHECK(hostToDevRedOp(&opFull, op, dataType, comm));