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

removed gfx940 and gfx941 (#1606)

Browse Source 
* removed gfx940 and gfx941

* removed gfx940 and gfx941

* Update "gfx94" to "gfx942" in init.cc

* Updated remaining "gfx94" updates to "gfx942"

* Update filenames and variables from gfx940 to gfx942

---------

Co-authored-by: akolliasAMD <akollias@amd.com>
This commit is contained in:
corey-derochie-amd
2025-03-20 09:34:53 -06:00
committed by GitHub
parent 6403afff4a
commit 6505639cf4
30 changed files with 74 additions and 79 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/init.cc
+4 -4
View File
@@ -97,7 +97,7 @@ static uint64_t hashUniqueId(ncclUniqueId const &id) {
ncclResult_t commSetUnrollFactor(struct ncclComm* comm) {
hipDeviceProp_t devProp;
CUDACHECK(hipGetDeviceProperties(&devProp, comm->cudaDev));
if(IsArchMatch(devProp.gcnArchName, "gfx908") || ((IsArchMatch(devProp.gcnArchName, "gfx94") || IsArchMatch(devProp.gcnArchName, "gfx950"))
if(IsArchMatch(devProp.gcnArchName, "gfx908") || ((IsArchMatch(devProp.gcnArchName, "gfx942") || IsArchMatch(devProp.gcnArchName, "gfx950"))
&& devProp.multiProcessorCount > 80))
comm->unroll = NCCL_UNROLL_2;
else
@@ -1338,7 +1338,7 @@ static ncclResult_t initTransportsRank(struct ncclComm* comm, struct ncclComm* p
allGather3Data[rank].nc = std::max(allGather3Data[rank].nc, 4/ringGraph->nChannels);
if (ringGraph->nChannels > MAXCHANNELS/2)
allGather3Data[rank].nc = 1;
if (IsArchMatch(comm->topo->nodes[GPU].nodes[idx].gpu.gcn, "gfx94")) {
if (IsArchMatch(comm->topo->nodes[GPU].nodes[idx].gpu.gcn, "gfx942")) {
// Multi-node MI300A
int managed = 0;
CUDACHECK(hipDeviceGetAttribute(&managed, hipDeviceAttributeDirectManagedMemAccessFromHost, 0));
@@ -1873,7 +1873,7 @@ static ncclResult_t ncclCommInitRankFunc(struct ncclAsyncJob* job_) {
// a CUDA memory reconfig on load (c.f. NVSHMEM issue)
#ifdef USE_INDIRECT_FUNCTION_CALL
CUDACHECK(hipGetDeviceProperties(&devProp, 0));
if (ncclParamSetStackSize() == 1 && !IsArchMatch(devProp.gcnArchName,"gfx94") && !IsArchMatch(devProp.gcnArchName,"gfx950")) {
if (ncclParamSetStackSize() == 1 && !IsArchMatch(devProp.gcnArchName,"gfx942") && !IsArchMatch(devProp.gcnArchName,"gfx950")) {
stackSize = rcclParamStackSizeOverride() ? rcclParamStackSizeOverride() : maxLocalSizeBytes;
if (stackSize == 0) {
if (IsArchMatch(devProp.gcnArchName,"gfx906"))
@@ -1939,7 +1939,7 @@ static ncclResult_t ncclCommInitRankFunc(struct ncclAsyncJob* job_) {
if (mscclEnabled() && (comm->topo->mscclEnabled || mscclForceEnabled()) && mscclppCommCompatible(comm)) {
hipDeviceProp_t devProp;
CUDACHECK(hipGetDeviceProperties(&devProp, cudaDev));
comm->mscclppCompatible = IsArchMatch(devProp.gcnArchName, "gfx94") || IsArchMatch(devProp.gcnArchName, "gfx950");
comm->mscclppCompatible = IsArchMatch(devProp.gcnArchName, "gfx942") || IsArchMatch(devProp.gcnArchName, "gfx950");
if (comm->mscclppCompatible) {
bool mapContainsId = (mscclpp_uniqueIdMap.count(job->commId) > 0);
auto& mscclppUniqueId = mscclpp_uniqueIdMap[job->commId];