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

Add Alltoallv RCCL kernel implementation (#269)

Browse Source 
* Add alltoallv API and implementation

* Extend Rome P2P channel limit to multinode and alltoall kernels

* topo_expl: fix compilation and sync up with main

* gtest: use RCCL alltoallv API

* Code review changes
This commit is contained in:
Wenkai Du
2020-09-30 16:25:36 -07:00
committed by GitHub
parent aa985bfb7e
commit b871ea3c0c
22 changed files with 300 additions and 60 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/graph/connect.cc
+1 -1
View File
@@ -291,7 +291,7 @@ ncclResult_t ncclTopoPostset(struct ncclComm* comm, int* firstRanks, struct nccl
int nc = nChannels*2;
if (gcn == 908) nc = std::max(nc, 4);
if (comm->topo->nodes[NET].count == 0 && comm->topo->type == RCCL_TOPO_CR8G) nc = nChannels*3;
if (comm->topo->nodes[NET].count == 0 && comm->topo->type == RCCL_TOPO_CR8G) nc = nChannels*4;
if (comm->topo->nodes[NET].count && comm->topo->type == RCCL_TOPO_4P2H_ROME) nc = 4*comm->topo->nodes[NET].count;
int end = std::min((int)ncclMaxNchannels(), std::max(nc, ncclMinNchannels()));