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

Gcn arch name (#886)

Browse Source 
We use CMake to determine if we're compiling against a version of ROCm that supports gcnArchName and handles architecture checking appropriately. It includes a few helper functions as drop ins for the functionality we used gcnArch for before; sometimes to enable flags, and sometimes to set frequencies.
This commit is contained in:
Audrey MP
2023-09-12 15:34:40 -04:00
committed by GitHub
parent e1dc4d5e42
commit e58ec78d35
16 changed files with 193 additions and 57 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/transport/net.cc
+1 -1
View File
@@ -190,7 +190,7 @@ static ncclResult_t sendSetup(struct ncclComm* comm, struct ncclTopoGraph* graph
if (req.netDev < 0) NCCLCHECK(ncclTopoGetNetDev(comm, myInfo->rank, graph, channelId, peerInfo->rank, &req.netDev, &proxyRank));
NCCLCHECK(ncclTopoCheckGdr(comm->topo, myInfo->busId, req.netDev, 1, &req.useGdr));
send->conn.flags |= req.useGdr ? NCCL_DIRECT_NIC : 0;
if (req.useGdr && comm->topo->nodes[GPU].nodes[0].gpu.gcn != 910 && comm->topo->nodes[GPU].nodes[0].gpu.gcn/10 != 94) {
if (req.useGdr && !IsArchMatch(comm->topo->nodes[GPU].nodes[0].gpu.gcn, "gfx90a") && !IsArchMatch(comm->topo->nodes[GPU].nodes[0].gpu.gcn, "gfx94")) {
CUDACHECK(hipDeviceGetAttribute((int*)&req.curr_hdp_reg, hipDeviceAttributeHdpMemFlushCntl, myInfo->cudaDev));
send->conn.curr_hdp_reg = req.curr_hdp_reg;
}
src/transport/p2p.cc
+1 -1
View File
@@ -354,7 +354,7 @@ ncclResult_t p2pSendSetup(struct ncclComm* comm, struct ncclTopoGraph* graph, st
INFO(NCCL_INIT|NCCL_P2P,"Ring %02d : %d -> %d failed to get link type and hop count", channelId, myInfo->rank, peerInfo->rank);
return ncclInternalError;
}
if (!isXGMI && comm->topo->nodes[GPU].nodes[0].gpu.gcn != 910 && comm->topo->nodes[GPU].nodes[0].gpu.gcn/10 != 94) {
if (!isXGMI && !IsArchMatch(comm->topo->nodes[GPU].nodes[0].gpu.gcn, "gfx90a") && !IsArchMatch(comm->topo->nodes[GPU].nodes[0].gpu.gcn, "gfx94")) {
CUDACHECK(hipDeviceGetAttribute((int*)&resources->next_hdp_reg, hipDeviceAttributeHdpMemFlushCntl,peerInfo->cudaDev));
TRACE(NCCL_INIT|NCCL_P2P,"Ring %02d : %d -> %d HDP %p", channelId, myInfo->rank, peerInfo->rank, resources->next_hdp_reg);
}