npkit: add broadcast trace (#1166)

[ROCm/rccl commit: a0cef69110]
This commit is contained in:
Wenkai Du
2024-05-07 14:00:16 -07:00
zatwierdzone przez GitHub
rodzic df3d462dd9
commit 0ff5fc0bad
3 zmienionych plików z 17 dodań i 0 usunięć
+13
Wyświetl plik
@@ -46,6 +46,13 @@ namespace {
}
#endif
#if defined(ENABLE_NPKIT) && defined(ENABLE_NPKIT_EVENT_BROADCAST_RING_ENTRY)
if (tid == 0) {
NpKit::CollectGpuEvent(NPKIT_EVENT_BROADCAST_RING_ENTRY, args->count*sizeof(T), 0, NPKIT_GET_GPU_TIMESTAMP(),
ncclShmem.comm.npKitEventCollectContexts + npKitCtxIdx);
}
#endif
T *inputBuf = (T*)args->sendbuff;
T *outputBuf = (T*)args->recvbuff;
Primitives<T, RedOp, FanSymmetric<1>, 0, Proto, 0>
@@ -73,6 +80,12 @@ namespace {
prims.recvCopySend(offset, nelem);
}
}
#if defined(ENABLE_NPKIT) && defined(ENABLE_NPKIT_EVENT_BROADCAST_RING_EXIT)
if (tid == 0) {
NpKit::CollectGpuEvent(NPKIT_EVENT_BROADCAST_RING_EXIT, args->count*sizeof(T), 0, NPKIT_GET_GPU_TIMESTAMP(),
ncclShmem.comm.npKitEventCollectContexts + npKitCtxIdx);
}
#endif
}
}