AI/rocm-systems
2
0
Fork 0
Koodi Ongelmat Pull-pyynnöt Actions Paketit Projektit Julkaisut Wiki Toiminta

Move hipify to cmake stage

Selaa lähdekoodia 
Add minimal ROCm/HIP version requirements for Graph support
This commit is contained in:
Wenkai Du
2022-11-07 14:09:26 -08:00
vanhempi 94ad7f6f51
commit 562dd87036
33 muutettua tiedostoa jossa 402 lisäystä ja 285 poistoa
Download Patch File Download Diff File Expand all files Collapse all files
src/group.cc
+6 -6
Näytä tiedosto
@@ -117,7 +117,7 @@ struct ncclPreconnectJob {
ncclResult_t ncclPreconnectFunc(struct ncclAsyncJob* job_) {
struct ncclPreconnectJob* job = (struct ncclPreconnectJob*)job_;
struct ncclComm* comm = job->comm;
CUDACHECK(hipSetDevice(comm->cudaDev));
CUDACHECK(cudaSetDevice(comm->cudaDev));
if (CPU_COUNT(&comm->cpuAffinity)) sched_setaffinity(0, sizeof(cpu_set_t), &comm->cpuAffinity);
NCCLCHECK(ncclTransportP2pSetup(comm, NULL, 1));
if (comm->p2pNet) NCCLCHECK(ncclTransportP2pSetup(comm, NULL, NCCL_CONN_IDX_P2P_NET));
@@ -138,7 +138,7 @@ static ncclResult_t doLaunches(struct ncclComm* head) {
bool capturingYes = false, capturingNo = false;
do {
(ncclCudaGraphValid(comm->tasks.capturingGraph) ? capturingYes : capturingNo) = true;
CUDACHECKGOTO(hipSetDevice(comm->cudaDev), result, failure);
CUDACHECKGOTO(cudaSetDevice(comm->cudaDev), result, failure);
NCCLCHECKGOTO(ncclLaunchPrepare(comm), result, failure);
if (useBarrier) ncclCommIntraBarrierIn(comm, 1);
comm = comm->groupNext;
@@ -170,7 +170,7 @@ static ncclResult_t doLaunches(struct ncclComm* head) {
struct ncclKernelPlan* plan = comm->unlaunchedPlansHead;
if (plan != nullptr) {
comm->unlaunchedPlansHead = plan->next;
CUDACHECKGOTO(hipSetDevice(comm->cudaDev), result, failure);
CUDACHECKGOTO(cudaSetDevice(comm->cudaDev), result, failure);
NCCLCHECKGOTO(ncclLaunchKernelBefore_NoUncapturedCuda(comm, plan), result, failure);
NCCLCHECKGOTO(ncclLaunchKernel(comm, plan), result, failure);
}
@@ -180,7 +180,7 @@ static ncclResult_t doLaunches(struct ncclComm* head) {
NCCLCHECKGOTO(ncclLaunchKernelAfter_NoCuda(comm, plan), result, failure);
}
} else { // Final round.
CUDACHECKGOTO(hipSetDevice(comm->cudaDev), result, failure);
CUDACHECKGOTO(cudaSetDevice(comm->cudaDev), result, failure);
NCCLCHECKGOTO(ncclLaunchFinish(comm), result, failure);
}
comm = next;
@@ -276,7 +276,7 @@ static ncclResult_t groupLaunch(struct ncclAsyncJob *job_) {
struct ncclIntruQueue<struct ncclAsyncJob, &ncclAsyncJob::next> *asyncJobsMain = gjob->asyncJobsPtr;
volatile bool *groupAbortFlag = gjob->abortFlagPtr;
CUDACHECKGOTO(hipGetDevice(&savedDev), ret, fail);
CUDACHECKGOTO(cudaGetDevice(&savedDev), ret, fail);
if (groupCommPreconnectHeadMain != nullptr) {
struct ncclComm* comm = groupCommPreconnectHeadMain;
@@ -368,7 +368,7 @@ static ncclResult_t groupLaunch(struct ncclAsyncJob *job_) {
*gjob->groupCommHeadPtr = nullptr;
*gjob->groupCommPreconnectHeadPtr = nullptr;
CUDACHECK(hipSetDevice(savedDev));
CUDACHECK(cudaSetDevice(savedDev));
exit:
return ret;