A warp whose even and odd lanes take different branches

A single warp (32 threads, contiguous threadIdx.x 0-31) executes this kernel. heavyA() and heavyB() each take roughly the same, nontrivial amount of time. What is the most accurate prediction of this warp's execution time compared to if all 32 threads took the same branch?

kernel.cu
__global__ void diverge(float* out) {
    int tid = threadIdx.x;
    if (tid % 2 == 0) {
        out[tid] = heavyA(tid);
    } else {
        out[tid] = heavyB(tid);
    }
}

3 of 46 in GPU / CUDA