nvprof produces unexpected branch efficiency results

Question

I followed the examples (the following codes) of warp divergence on the textbook "Professional CUDA C Programming".

__global__ void math_kernel1(float *c) {
    int tid = blockIdx.x * blockDim.x + threadIdx.x;
    float a, b;
    a = b = 0.f;
    if (tid % 2 == 0) {
        a = 100.0f;
    } else {
        b = 200.0f;
    }
    c[tid] = a + b;
}

__global__ void math_kernel2(float *c) {
    int tid = blockIdx.x * blockDim.x + threadIdx.x;
    float a, b;
    a = b = 0.f;
    if ((tid / warpSize) % 2 == 0) {
        a = 100.0f;
    } else {
        b = 200.0f;
    }
    c[tid] = a + b;
}

__global__ void math_kernel3(float *c) {
    int tid = blockIdx.x * blockDim.x + threadIdx.x;
    float a, b;
    bool ipred = (tid % 2 == 0);
    a = b = 0.f;
    if (ipred) {
        a = 100.0f;
    }
    if (!ipred) {
        b = 200.0f;
    }

    c[tid] = a + b;
}

Obviously (and written on the textbook), the math_kernel2 should have best branch efficiency, math_kernel1 follows and math_kernel3 has the worst result. However, the nvprof report gives me results that are contradictory to the textbook. I benchmarked these kernels on GTX 1080 Ti with CUDA 8.0 (I also added the compiler flags -g -G to nvcc to disable optimizations) and it reports the following branch efficiency:

1. math_kernel1 83.33%
2. math_kernel2 100.00%
3. math_kernel3 100.00% (expected to be less than math_kernel1 and it is 71.43% on the textbook)