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

Revert "[SRC] Enable unroll=1 for gfx950 (#1602)" (#1667)

Browse Source 
* Revert "[SRC] Enable unroll=1 for gfx950 (#1602)"
This reverts commit 210f90ae0f.

* Update Changelog

---------

Signed-off-by: nileshnegi <Nilesh.Negi@amd.com>

[ROCm/rccl commit: 329e13efff]
This commit is contained in:
Nilesh M Negi
2025-04-30 23:33:08 -05:00
committed by GitHub
parent 5580cb7574
commit a6972c0d09
9 changed files with 67 additions and 125 deletions
Download Patch File Download Diff File Expand all files Collapse all files
projects/rccl/src/init.cc
+2 -3
View File
@@ -99,9 +99,8 @@ static uint64_t hashUniqueId(ncclUniqueId const &id) {
ncclResult_t commSetUnrollFactor(struct ncclComm* comm) {
hipDeviceProp_t devProp;
CUDACHECK(hipGetDeviceProperties(&devProp, comm->cudaDev));
if(IsArchMatch(devProp.gcnArchName, "gfx950"))
comm->unroll = NCCL_UNROLL_1;
else if(IsArchMatch(devProp.gcnArchName, "gfx908") || ((IsArchMatch(devProp.gcnArchName, "gfx942") && devProp.multiProcessorCount > 80)))
if(IsArchMatch(devProp.gcnArchName, "gfx908") || ((IsArchMatch(devProp.gcnArchName, "gfx942") || IsArchMatch(devProp.gcnArchName, "gfx950"))
&& devProp.multiProcessorCount > 80))
comm->unroll = NCCL_UNROLL_2;
else
comm->unroll = NCCL_UNROLL_4;