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

Merge remote-tracking branch 'nccl/master' into develop

Browse Source 
[ROCm/rccl commit: a79d9e3586]
This commit is contained in:
Wenkai Du
2022-09-09 01:20:52 +00:00
parent fe99249cde c9a50a9ec5
commit 7874a99c75
90 changed files with 5517 additions and 3115 deletions
Download Patch File Download Diff File Expand all files Collapse all files
projects/rccl/src/graph/paths.cc
+12 -13
View File
@@ -448,10 +448,10 @@ NCCL_PARAM(PxnDisable, "PXN_DISABLE", 1);
// Net v4 plugins don't have non-blocking connect/accept. We can't therefore use
// remote proxies without risking deadlocks
int ncclPxnDisable() {
int ncclPxnDisable(struct ncclComm* comm) {
static int pxnDisable = -1;
if (pxnDisable == -1) {
if (ncclNetVersion() == 4) {
if (comm && ncclNetVersion(comm) == 4) {
INFO(NCCL_INIT, "PXN Disabled as plugin is v4");
pxnDisable = 1;
} else {
@@ -490,7 +490,7 @@ ncclResult_t ncclTopoGetPxnRanks(struct ncclComm* comm, int** intermediateRanks,
return ncclSuccess;
}
ncclResult_t ncclTopoComputePaths(struct ncclTopoSystem* system, struct ncclPeerInfo* peerInfos) {
ncclResult_t ncclTopoComputePaths(struct ncclTopoSystem* system, struct ncclComm* comm) {
// Precompute paths between GPUs/NICs.
// Remove everything in case we're re-computing
@@ -518,16 +518,16 @@ ncclResult_t ncclTopoComputePaths(struct ncclTopoSystem* system, struct ncclPeer
}
}
if (peerInfos == NULL) continue;
if (comm == NULL) continue;
// Remove GPUs we can't talk to because of containers.
struct ncclPeerInfo* dstInfo = peerInfos+system->nodes[GPU].nodes[g].gpu.rank[0];
struct ncclPeerInfo* dstInfo = comm->peerInfo+system->nodes[GPU].nodes[g].gpu.rank[0];
for (int p=0; p<system->nodes[GPU].count; p++) {
if (p == g) continue;
struct ncclPeerInfo* srcInfo = peerInfos+system->nodes[GPU].nodes[p].gpu.rank[0];
struct ncclPeerInfo* srcInfo = comm->peerInfo+system->nodes[GPU].nodes[p].gpu.rank[0];
int shm;
NCCLCHECK(ncclTransports[TRANSPORT_SHM].canConnect(&shm, system, NULL, srcInfo, dstInfo));
NCCLCHECK(ncclTransports[TRANSPORT_SHM]->canConnect(&shm, system, NULL, srcInfo, dstInfo));
int p2p;
NCCLCHECK(ncclTransports[TRANSPORT_P2P].canConnect(&p2p, system, NULL, srcInfo, dstInfo));
NCCLCHECK(ncclTransports[TRANSPORT_P2P]->canConnect(&p2p, system, NULL, srcInfo, dstInfo));
if (shm == 0 && p2p == 0) {
// Mark this peer as inaccessible. We'll trim it later.
system->nodes[GPU].nodes[p].paths[GPU][g].count = 0;
@@ -543,7 +543,7 @@ ncclResult_t ncclTopoComputePaths(struct ncclTopoSystem* system, struct ncclPeer
for (int g=0; g<system->nodes[GPU].count; g++) {
// Check whether we can access the NIC through another NVLink-connected GPU (PXN)
struct ncclTopoNode* gpu = system->nodes[GPU].nodes+g;
if (ncclPxnDisable() != 1 && gpu->paths[NET][n].type > PATH_PXB) {
if (ncclPxnDisable(comm) != 1 && gpu->paths[NET][n].type > PATH_PXB) {
int pxnGpu = -1;
for (int p=0; p<system->nodes[GPU].count; p++) {
@@ -556,7 +556,6 @@ ncclResult_t ncclTopoComputePaths(struct ncclTopoSystem* system, struct ncclPeer
pxnGpu = p;
int netDev;
NCCLCHECK(ncclTopoGetLocalNet(system, peerNode->gpu.rank[0], &netDev));
// To ensure proper balancing, use preferably a local GPU which advertised that NIC as its preferred one.
if (netDev == netNode->id) break;
@@ -602,8 +601,8 @@ ncclResult_t ncclTopoTrimSystem(struct ncclTopoSystem* system, struct ncclComm*
}
for (int j=0; j<gpu->gpu.nRanksPerGpu; j++ ) {
if (gpu->gpu.rank[j] == comm->rank) {
myDomain = domains[g];
break;
myDomain = domains[g];
break;
}
}
}
@@ -768,7 +767,7 @@ ncclResult_t ncclTopoComputeP2pChannels(struct ncclComm* comm) {
// We want to spread channels used when there aren't many and progressively
// fill the whole space of nChannels. To do so we mirror the bits in the
// nChannels space.
for (int c=0; c<comm->p2pnChannelsPerPeer; c++) {
for (int c=0; c<comm->p2pnChannels; c++) {
int mirror = 0;
for (int b=1, mb=(comm->p2pnChannels>>1); b<comm->p2pnChannels; b<<=1, mb>>=1) if (c & b) mirror |= mb;
comm->p2pChannels[c] = mirror;
projects/rccl/src/graph/search.cc
+8 -4
View File
@@ -275,8 +275,8 @@ ncclResult_t ncclTopoReplayGetGpu(struct ncclTopoSystem* system, struct ncclTopo
for (int i=0; i<ngpus; i++) {
for (int j=0; j<system->nodes[GPU].nodes[i].gpu.nRanksPerGpu; j++ ) {
if (system->nodes[GPU].nodes[i].gpu.rank[j] == nextRank) {
*g = i;
return ncclSuccess;
*g = i;
return ncclSuccess;
}
}
}
@@ -1103,10 +1103,14 @@ ncclResult_t ncclTopoGetNetDev(struct ncclComm* comm, int rank, struct ncclTopoG
NCCLCHECK(ncclTopoGetLocalNet(comm->topo, rank, dev));
*proxyRank = rank;
int pxnLevel = ncclPxnDisable() == 1 ? 0 : ncclParamP2pPxnLevel();
int pxnLevel = ncclPxnDisable(comm) == 1 ? 0 : ncclParamP2pPxnLevel();
// See whether we can use the remote rank preferred device.
if (ncclParamCrossNic() == 0 || (pxnLevel != 0)) {
int netDev = comm->peerInfo[peerRank].netDev;
// Find local NIC number close to local cudaDev
int cudaDev = comm->peerInfo[peerRank].cudaDev;
int localRank;
if (ncclTopoDevToRank(comm->topo, cudaDev, &localRank) != ncclSuccess) return ncclSuccess;
int netDev = comm->peerInfo[localRank].netDev;
int n;
// Check that device exists on our node
if (ncclParamCrossNic() == 0) {
projects/rccl/src/graph/topo.cc
+13 -9
View File
@@ -724,11 +724,11 @@ ncclResult_t ncclTopoGetSystem(struct ncclComm* comm, struct ncclTopoSystem** sy
// Auto-detect NICs if needed. net/collnet share the same xml/graph nodes,
// so we start with collnet so that it has precedence.
int netDevCount = 0;
if (collNetSupport()) {
NCCLCHECK(collNetDevices(&netDevCount));
if (collNetSupport(comm)) {
NCCLCHECK(collNetDevices(comm, &netDevCount));
for (int n=0; n<netDevCount; n++) {
ncclNetProperties_t props;
NCCLCHECK(collNetGetProperties(n, &props));
NCCLCHECK(collNetGetProperties(comm, n, &props));
struct ncclXmlNode* netNode;
NCCLCHECK(ncclTopoFillNet(xml, props.pciPath, props.name, &netNode));
NCCLCHECK(xmlSetAttrInt(netNode, "keep", 1));
@@ -737,16 +737,18 @@ ncclResult_t ncclTopoGetSystem(struct ncclComm* comm, struct ncclTopoSystem** sy
NCCLCHECK(xmlInitAttrInt(netNode, "port", props.port));
NCCLCHECK(xmlInitAttrUint64(netNode, "guid", props.guid));
NCCLCHECK(xmlInitAttrInt(netNode, "maxconn", props.maxComms));
NCCLCHECK(xmlInitAttrInt(netNode, "gdr", props.ptrSupport & NCCL_PTR_CUDA ? 1 : 0));
bool gdrSupport = (props.ptrSupport & NCCL_PTR_CUDA) || (comm->dmaBufSupport && (props.ptrSupport & NCCL_PTR_DMABUF));
INFO(NCCL_NET,"NET/%s : GPU Direct RDMA %s for HCA %d '%s'", comm->ncclNet->name, gdrSupport ? "Enabled" : "Disabled", n, props.name);
NCCLCHECK(xmlInitAttrInt(netNode, "gdr", gdrSupport));
NCCLCHECK(xmlInitAttrInt(netNode, "coll", 1));
}
}
if (netDevCount == 0) {
NCCLCHECK(ncclNetDevices(&netDevCount));
NCCLCHECK(ncclNetDevices(comm, &netDevCount));
}
for (int n=0; n<netDevCount; n++) {
ncclNetProperties_t props;
NCCLCHECK(ncclNetGetProperties(n, &props));
NCCLCHECK(ncclNetGetProperties(comm, n, &props));
struct ncclXmlNode* netNode;
NCCLCHECK(ncclTopoFillNet(xml, props.pciPath, props.name, &netNode));
NCCLCHECK(xmlSetAttrInt(netNode, "keep", 1));
@@ -756,7 +758,9 @@ ncclResult_t ncclTopoGetSystem(struct ncclComm* comm, struct ncclTopoSystem** sy
NCCLCHECK(xmlInitAttrFloat(netNode, "latency", props.latency));
NCCLCHECK(xmlInitAttrUint64(netNode, "guid", props.guid));
NCCLCHECK(xmlInitAttrInt(netNode, "maxconn", props.maxComms));
NCCLCHECK(xmlInitAttrInt(netNode, "gdr", props.ptrSupport & NCCL_PTR_CUDA ? 1 : 0));
bool gdrSupport = (props.ptrSupport & NCCL_PTR_CUDA) || (comm->dmaBufSupport && (props.ptrSupport & NCCL_PTR_DMABUF));
INFO(NCCL_NET,"NET/%s : GPU Direct RDMA %s for HCA %d '%s'", comm->ncclNet->name, gdrSupport ? "Enabled" : "Disabled", n, props.name);
NCCLCHECK(xmlInitAttrInt(netNode, "gdr", gdrSupport));
}
// Remove XML branches which don't have a node with keep="1" (typically when importing a topology)
@@ -903,8 +907,8 @@ ncclResult_t ncclTopoGetLocalRank(struct ncclTopoSystem* system, int rank, int*
for (int g=0; g<system->nodes[GPU].count; g++) {
for ( int j=0; j<system->nodes[GPU].nodes[g].gpu.nRanksPerGpu; j++ ){
if (system->nodes[GPU].nodes[g].gpu.rank[j] == rank) {
*localRank = g;
return ncclSuccess;
*localRank = g;
return ncclSuccess;
}
}
}
projects/rccl/src/graph/topo.h
+14 -3
View File
@@ -198,20 +198,31 @@ static ncclResult_t ncclTopoRankToIndex(struct ncclTopoSystem* system, int rank,
for (int i=0; i<system->nodes[GPU].count; i++) {
for (int j=0; j<system->nodes[GPU].nodes[i].gpu.nRanksPerGpu; j++ ) {
if (system->nodes[GPU].nodes[i].gpu.rank[j] == rank) {
*index = i;
return ncclSuccess;
*index = i;
return ncclSuccess;
}
}
}
return ncclInternalError;
}
static ncclResult_t ncclTopoDevToRank(struct ncclTopoSystem* system, int dev, int* rank) {
*rank = -1;
for (int i=0; i<system->nodes[GPU].count; i++) {
if (system->nodes[GPU].nodes[i].gpu.dev == dev) {
*rank = system->nodes[GPU].nodes[i].gpu.rank[0];
return ncclSuccess;
}
}
return ncclInternalError;
}
// Returns XGMI speed in GB/s
static float ncclTopoXGMISpeed(int gcn) {
return gcn == 910 ? MI200_XGMI_WIDTH : VEGA_XGMI_WIDTH;
}
#define ncclGetKernelIndex(p_comm) \
(((p_comm)->topo->ll128Enabled ? 1 : 0)*2 + ((p_comm)->hostDevComm.collTraceThread ? 1 : 0))
(((p_comm)->topo->ll128Enabled ? 1 : 0)*2 + ((p_comm)->collTraceThread ? 1 : 0))
#endif
projects/rccl/src/graph/tuning.cc
+3 -3
View File
@@ -235,11 +235,11 @@ ncclResult_t ncclTopoTuneModel(struct ncclComm* comm, int minCompCap, int maxCom
int simpleDefaultThreads = (ringGraph->speedIntra*ringGraph->nChannels <= PCI_WIDTH) ? 256 : NCCL_SIMPLE_MAX_NTHREADS;
comm->maxThreads[NCCL_ALGO_RING][NCCL_PROTO_SIMPLE] =
#if defined(__HIP_PLATFORM_HCC__) || defined(__HCC__) || defined(__HIPCC__)
getNthreads("NCCL_NTHREADS", ncclParamNthreads(), 4*WARP_SIZE, NCCL_MAX_NTHREADS, simpleDefaultThreads);
getNthreads("NCCL_NTHREADS", ncclParamNthreads(), 4*comm->WarpSize, NCCL_MAX_NTHREADS, simpleDefaultThreads);
comm->maxThreads[NCCL_ALGO_TREE][NCCL_PROTO_SIMPLE] = comm->maxThreads[NCCL_ALGO_COLLNET][NCCL_PROTO_SIMPLE] =
getNthreads("NCCL_NTHREADS", ncclParamNthreads(), 4*WARP_SIZE, NCCL_MAX_NTHREADS, NCCL_MAX_NTHREADS);
getNthreads("NCCL_NTHREADS", ncclParamNthreads(), 4*comm->WarpSize, NCCL_MAX_NTHREADS, NCCL_MAX_NTHREADS);
comm->maxThreads[NCCL_ALGO_RING][NCCL_PROTO_LL] = comm->maxThreads[NCCL_ALGO_TREE][NCCL_PROTO_LL] = comm->maxThreads[NCCL_ALGO_COLLNET][NCCL_PROTO_LL] =
getNthreads("NCCL_NTHREADS", ncclParamNthreads(), 4*WARP_SIZE, NCCL_MAX_NTHREADS, NCCL_MAX_NTHREADS);
getNthreads("NCCL_NTHREADS", ncclParamNthreads(), 4*comm->WarpSize, NCCL_MAX_NTHREADS, NCCL_MAX_NTHREADS);
#else
getNthreads("NCCL_NTHREADS", ncclParamNthreads(), 2*WARP_SIZE, NCCL_SIMPLE_MAX_NTHREADS, simpleDefaultThreads);
comm->maxThreads[NCCL_ALGO_TREE][NCCL_PROTO_SIMPLE] =