def use_shfl_sync_with_val(ary, into): i = cuda.grid(1) val = cuda.shfl_sync(0xffffffff, into, 0) ary[i] = val
def use_shfl_sync_idx(ary, idx): i = cuda.grid(1) val = cuda.shfl_sync(0xffffffff, i, idx) ary[i] = val
def use_shfl_sync_with_val(ary, into): i = cuda.grid(1) val = cuda.shfl_sync(0xffffffff, into, 0) ary[i] = val
def use_shfl_sync_idx(ary, idx): i = cuda.grid(1) val = cuda.shfl_sync(0xffffffff, i, idx) ary[i] = val