npkit: add broadcast trace (#1166)

src/device/broadcast.h

@@ -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
   }
 }