npkit: add broadcast trace (#1166)

[ROCm/rccl commit: a0cef69110]
Tento commit je obsažen v:
Wenkai Du
2024-05-07 14:00:16 -07:00
odevzdal GitHub
rodič df3d462dd9
revize 0ff5fc0bad
3 změnil soubory, kde provedl 17 přidání a 0 odebrání
+2
Zobrazit soubor
@@ -340,6 +340,8 @@ if ($npkit_enabled); then
-DENABLE_NPKIT_EVENT_MSCCL_RECV_REDUCE_COPY_EXIT \
-DENABLE_NPKIT_EVENT_MSCCL_INIT_ENTRY \
-DENABLE_NPKIT_EVENT_MSCCL_INIT_EXIT \
-DENABLE_NPKIT_EVENT_BROADCAST_RING_ENTRY \
-DENABLE_NPKIT_EVENT_BROADCAST_RING_EXIT \
-DENABLE_NPKIT_PRIM_COLLECT_DATA_PROCESS_TIME"
fi
+13
Zobrazit soubor
@@ -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
}
}
+2
Zobrazit soubor
@@ -127,4 +127,6 @@
#define NPKIT_EVENT_MSCCL_INIT_ENTRY 0x66
#define NPKIT_EVENT_MSCCL_INIT_EXIT 0x67
#define NPKIT_EVENT_BROADCAST_RING_ENTRY 0x68
#define NPKIT_EVENT_BROADCAST_RING_EXIT 0x69
#endif