AI/rocm-systems
2
0
Çatalla 0
Kod Konular Değişiklik İstekleri İşlemler Paketler Projeler Sürüm Viki Aktivite

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

Kaynağa Gözat 
* 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>
Bu işleme şunda yer alıyor:
Mustafa Abduljabbar
2025-08-26 15:03:54 -04:00
işlemeyi yapan: GitHub
ebeveyn b882af9ffd
işleme 277747c199
18 değiştirilmiş dosya ile 286 ekleme ve 170 silme
Yama Dosyasını İndir Diff Dosyasını İndir Tüm dosyaları genişlet Tüm dosyaları daralt
cmake/scripts/add_unroll.sh
Normal dosya → Çalıştırılabilir dosya
+13 -4
Dosyayı Görüntüle
@@ -21,13 +21,22 @@
HIP_FILE=$1
if [[ "$HIP_FILE" =~ .*/src/device/.*\.h ]]; then
perl -pi -e 's/(template<typename T, typename RedOp(?:, typename Proto)?)(, bool isNetOffload.*?)?>/\1, int USE_ACC, int COLL_UNROLL\2>/g' "$HIP_FILE"
perl -pi -e 's/(template<typename T, typename RedOp(?:, typename Proto)?)(, bool isNetOffload.*?)?>/\1, int USE_ACC, int COLL_UNROLL, int Pipeline\2>/g' "$HIP_FILE"
perl -pi -e 's/(template<typename T, typename RedOp(?:, typename Proto)?(?:, int RCCLMetadata)?)(, bool isNetOffload.*?)?>/\1, int USE_ACC, int COLL_UNROLL, int Pipeline\2>/g' "$HIP_FILE"
perl -pi -e 's/(ProtoSimple<[^,]*?,[^,]+?)>/\1, USE_ACC, COLL_UNROLL>/g' "$HIP_FILE"
perl -pi -e 's/(runRing<T.*?)((, (true|false))?>\()/\1, USE_ACC, COLL_UNROLL\2/g' "$HIP_FILE"
perl -pi -e 's/(runTreeUpDown<T.*?)>\(/\1, USE_ACC, COLL_UNROLL>(/' "$HIP_FILE"
perl -pi -e 's/(runTreeSplit<T.*?)>\(/\1, USE_ACC, COLL_UNROLL>(/' "$HIP_FILE"
sed -i "s/\\(struct RunWorkColl<ncclFunc[^>]*\\)>*/\\1, USE_ACC, COLL_UNROLL>/" "$HIP_FILE"
sed -i "s/\\(struct RunWorkBatch<ncclFunc[^>]*\\)>*/\\1, USE_ACC, COLL_UNROLL>/" "$HIP_FILE"
echo "Added COLL_UNROLL and USE_ACC template arguments to $HIP_FILE"
perl -pi -e 's/(runTreeSplit<T, RedOp, (ProtoLL|ProtoLL128), USE_ACC, COLL_UNROLL.*?)>/\1, 0>/' "$HIP_FILE"
perl -pi -e 's/(runTreeUpDown<T, RedOp, (ProtoLL|ProtoLL128), USE_ACC, COLL_UNROLL.*?)>/\1, 0>/' "$HIP_FILE"
perl -pi -e 's/(runRing<T, RedOp, (ProtoLL|ProtoLL128), USE_ACC, COLL_UNROLL.*?)>/\1, 0>/' "$HIP_FILE"
perl -pi -e 's/(runRing<T, RedOp, (ProtoLL|ProtoLL128), (RCCL_ONE_NODE_RING_SIMPLE|RCCL_METADATA_EMPTY), USE_ACC, COLL_UNROLL.*?)>/\1, 0>/' "$HIP_FILE"
perl -pi -e 's/(runRing<T, RedOp, Proto, (RCCL_ONE_NODE_RING_SIMPLE|RCCL_METADATA_EMPTY), USE_ACC, COLL_UNROLL.*?)>/\1, Pipeline>/' "$HIP_FILE"
perl -pi -e 's/(runRing<T, RedOp, Proto, USE_ACC, COLL_UNROLL.*?)>/\1, Pipeline>/' "$HIP_FILE"
perl -pi -e 's/(runTreeSplit<T, RedOp, Proto, USE_ACC, COLL_UNROLL.*?)>/\1, Pipeline>/' "$HIP_FILE"
perl -pi -e 's/(runTreeUpDown<T, RedOp, Proto, USE_ACC, COLL_UNROLL.*?)>/\1, Pipeline>/' "$HIP_FILE"
sed -i "s/\\(struct RunWorkBatch<ncclFunc[^>]*\\)>*/\\1, USE_ACC, COLL_UNROLL, Pipeline>/" "$HIP_FILE"
sed -i "s/\\(RunWorkColl<[^,]*,[^,]*,[^,]*,[^,]*,[^>]*\\)>/\\1, USE_ACC, COLL_UNROLL, Pipeline>/" "$HIP_FILE"
fi