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

Select sendrecv path based on collective data size (#391)

Browse Source 
* Select sendrecv path based on collective data size

* Add comments on packing and unpacking group field

* Toggling RCCL_P2P_NET_DISABLE in combined calls unit tests
This commit is contained in:
Wenkai Du
2021-06-10 17:51:04 -07:00
committed by GitHub
parent f6f5e16fe6
commit 6dcae8a459
17 changed files with 102 additions and 45 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/collectives/device/all_reduce.h
+1 -1
View File
@@ -37,7 +37,7 @@ class ncclFunction<ncclFuncAllReduce, NCCL_ALGO_RING, NCCL_PROTO_SIMPLE, FUNC, T
T * __restrict__ thisOutput = (T*)args->recvbuff;
ncclPrimitives<UNROLL, ALLREDUCE_CHUNKSTEPS/ALLREDUCE_SLICESTEPS, ALLREDUCE_SLICESTEPS, T, 1, 1, 1, FUNC>
prims(tid, nthreads, &ring->prev, &ring->next, thisOutput, stepSize, channel, comm, ncclShmem->ptrs, args->coll.connIndex);
prims(tid, nthreads, &ring->prev, &ring->next, thisOutput, stepSize, channel, comm, ncclShmem->ptrs, PACK_GROUP(0, args->coll.connIndex));
for (ssize_t gridOffset = 0; gridOffset < size; gridOffset += nranks*loopSize) {
ssize_t realChunkSize = min(chunkSize, DIVUP(size-gridOffset,nranks*nChannels));