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

[Device] Add dynamic fetch/reduce pipelining for reduction collectives - Simple protocol (#1861)

Browse Source 
* Support pipelining codegen and template specialization

* Support ReduceCopy pipelining for AllReduce, ReduceScatter, and Reduce (currently enabled for bfloat16)

* Remove need for FUNC_INDEX_TOTAL

* Add pipeline field to device function key construction logic

* Avoid unneeded codegen for LL/LL64 kernels

* Modify conditions and add pipeline dtypes env

* Optimize selection for both gfx942 and gfx950

* Increase pipeline bitfield width

* Use __forceinline__ for all device functions

* Realign reduceCopy with original form

* Add opt-out option to enable perf debugs

* Remove force-reduce-pipelining option from README

* Update CHANGELOG.md

---------

Co-authored-by: Jeffrey Novotny <jnovotny@amd.com>
This commit is contained in:
Mustafa Abduljabbar
2025-08-26 15:03:54 -04:00
committed by GitHub
parent b882af9ffd
commit 277747c199
18 changed files with 286 additions and 170 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/include/comm.h
+1 -1
View File
@@ -207,7 +207,7 @@ struct ncclTaskColl {
size_t trafficBytes;
int32_t nMaxChannels:8;
int32_t nWarps:8;
int32_t algorithm:8, protocol:8;
int32_t algorithm:8, protocol:8, pipeline:8;
uint32_t isCollnet:1, isNvls:1;
uint32_t devFuncId:30;
int regBufType;