Add gather, scatter and alltoall collectives

Introducing 3 new APIs:
ncclResult_t  ncclGather(const void* sendbuff, void* recvbuff, size_t sendcount,
    ncclDataType_t datatype, int root, ncclComm_t comm, hipStream_t stream);
ncclResult_t  ncclScatter(const void* sendbuff, void* recvbuff,
    size_t recvcount, ncclDataType_t datatype, int root, ncclComm_t comm,
    hipStream_t stream);
ncclResult_t  ncclAllToAll(const void* sendbuff, void* recvbuff, size_t count,
    ncclDataType_t datatype, ncclComm_t comm, hipStream_t stream);

Only out of place operation is supported.
Preprocessor symbol RCCL_GATHER_SCATTER=1 indicates API availibility.
By default the APIs launche RCCL kernel implementation, which can be disabled by
RCCL_ALLTOALL_KERNEL_DISABLE=1. Then the APIs use wrapper around ncclSend and ncclRecv.
This commit is contained in:
Wenkai Du
2020-03-18 17:03:03 -07:00
parent 71ec3e09df
commit e80e29573c
24 changed files with 617 additions and 15 deletions
+3
View File
@@ -200,6 +200,9 @@ int main(int argc,char* argv[])
for (int i = 0; i < nranks; i++) {
comm[i].rank = i;
comm[i].nRanks = nranks;
comm[i].p2plist.count=0;
NCCLCHECK(ncclCalloc(&comm[i].p2plist.connect.recv, MAXCHANNELS*comm->nRanks));
NCCLCHECK(ncclCalloc(&comm[i].p2plist.connect.send, MAXCHANNELS*comm->nRanks));
node_model = network.GetNode(i);
assert(node_model!=0);
comm[i].topo = node_model->getSystem(i);