AI/rocm-systems
2
0
Vork 0
Code Kwesties Pull-aanvragen Actions Paketten Projecten Publicaties Wiki Activiteit

NPKit: separate time stamps for GPU access from different blocks (#1229)

Bladeren bron 
To avoid races in memory access in GPU
This commit is contained in:
Wenkai Du
2024-06-28 08:00:22 -07:00
gecommit door GitHub
bovenliggende b09ea29d66
commit 9d8f68b4ee
7 gewijzigde bestanden met toevoegingen van 17 en 21 verwijderingen
Download Patch-bestand Download Diff-bestand Expand all files Collapse all files
src/device/sendrecv.h
+2 -4
Bestand weergeven
@@ -26,8 +26,7 @@ struct RunWork<ncclFuncSendRecv, T, RedOp, NCCL_ALGO_RING, NCCL_PROTO_SIMPLE> {
#if defined(ENABLE_NPKIT) && defined(ENABLE_NPKIT_EVENT_TIME_SYNC_CPU)
if (isNpKitThread) {
uint64_t* cpuTimestamp = ncclShmem.comm.cpuTimestamp;
NpKit::CollectGpuEvent(NPKIT_EVENT_TIME_SYNC_CPU, 0, 0, *cpuTimestamp,
NpKit::CollectGpuEvent(NPKIT_EVENT_TIME_SYNC_CPU, 0, 0, NPKIT_GET_CPU_TIMESTAMP_FROM_BLOCK,
ncclShmem.comm.npKitEventCollectContexts + npKitCtxIdx);
}
#endif
@@ -127,8 +126,7 @@ struct RunWork<ncclFuncSendRecv, T, RedOp, NCCL_ALGO_RING, NCCL_PROTO_SIMPLE> {
#if defined(ENABLE_NPKIT) && defined(ENABLE_NPKIT_EVENT_TIME_SYNC_CPU)
if (isNpKitThread) {
uint64_t* cpuTimestamp = ncclShmem.comm.cpuTimestamp;
NpKit::CollectGpuEvent(NPKIT_EVENT_TIME_SYNC_CPU, 0, 0, *cpuTimestamp,
NpKit::CollectGpuEvent(NPKIT_EVENT_TIME_SYNC_CPU, 0, 0, NPKIT_GET_CPU_TIMESTAMP_FROM_BLOCK,
ncclShmem.comm.npKitEventCollectContexts + npKitCtxIdx);
}
#endif