Why NCCL LL128 proto need to load data twice?

[MARD1NO]

I notice the code below in prims_ll128.h

do {
          needReload = false;
          #pragma unroll
          for (int u=0; u<ELEMS_PER_THREAD; u+=2) {
            load128(ptr+u*WARP_SIZE, vr[u], vr[u+1]);
            needReload |= flagThread && (vr[u+1] != flag);
          }
          needReload &= (0 == checkAbort(spins, i, 0));
        } while (__any_sync(WARP_MASK, needReload));

#pragma unroll
for (int u=0; u<ELEMS_PER_THREAD; u+=2)
  load128(ptr+u*WARP_SIZE, vr[u], vr[u+1]);

I think when the first do while loop finish, the register is loaded remote data successfully, so why we need to load again?...