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

2.20.3-1

Browse Source 
Add support for alternating rings, allow for cross-nic rings without
cross-rail communication.
Add support for user buffer registration for network send/recv.
Optimize aggregated operations to better utilize all channels.
Add flattening for BCM PCI gen5 switches.
Add support for inter-node NVLink communication
Add support for port fusion in NET/IB.
Add support for ReduceScatter and AllGather using Collnet.
Update net API to v8.
Fix hang during A2A connection.
This commit is contained in:
Sylvain Jeaugey
2024-02-05 05:06:02 -08:00
parent b6d7438d31
commit b6475625fb
74 changed files with 4632 additions and 2165 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/device/sendrecv.h
+4 -4
View File
@@ -26,13 +26,13 @@ struct RunWork<ncclFuncSendRecv, T, RedOp, NCCL_ALGO_RING, NCCL_PROTO_SIMPLE> {
if (args->proto == NCCL_PROTO_LL) chunkSize /= 2;
int const peer = args->peer;
Primitives<T, RedOp, FanAsymmetric<0, 1>, 1, Proto, 1> prims
(tid, nthreads, nullptr, &peer, buff, nullptr, /*redOpArg(ignored)=*/0, group, 1, 1, nullptr, ncclShmem.comm.p2pChunkSize/sizeof(T));
(tid, nthreads, nullptr, &peer, buff, nullptr, /*redOpArg(ignored)=*/0, group, 1, 1, nullptr, args, ncclShmem.comm.p2pChunkSize/sizeof(T));
size_t offset = 0;
do {
int nelem = min(size_t(chunkSize), count-offset);
prims.directSend(offset, offset, nelem);
offset += nelem;
} while(offset < count);
} while(offset < count && args->reg == 0);
}
}
@@ -45,13 +45,13 @@ struct RunWork<ncclFuncSendRecv, T, RedOp, NCCL_ALGO_RING, NCCL_PROTO_SIMPLE> {
if (args->proto == NCCL_PROTO_LL) chunkSize /= 2; // This is to account for chunkEffectiveSize
int const peer = args->peer;
Primitives<T, RedOp, FanAsymmetric<1, 0>, 1, Proto, 1> prims
(tid, nthreads, &peer, nullptr, nullptr, buff, /*redOpArg(ignored)=*/0, group, 1, 1, nullptr, ncclShmem.comm.p2pChunkSize/sizeof(T));
(tid, nthreads, &peer, nullptr, nullptr, buff, /*redOpArg(ignored)=*/0, group, 1, 1, nullptr, args, ncclShmem.comm.p2pChunkSize/sizeof(T));
size_t offset = 0;
do {
int nelem = min(size_t(chunkSize), count-offset);
prims.directRecv(offset, nelem);
offset += nelem;
} while(offset < count);
} while(offset < count && args->reg == 0);
}
}