Reading one constant-memory address versus 32 different ones per warp

__constant__ float coeffs[256]; is populated once via cudaMemcpyToSymbol before launch. Kernel A has every thread in a warp read coeffs[0] (the same address). Kernel B has each thread read coeffs[threadIdx.x % 256] (32 different addresses). What's the performance difference, assuming both are otherwise identical and the constant cache is warm?

kernel.cu
__constant__ float coeffs[256];

// Kernel A
__global__ void useSameCoeff(float* out, const float* in) {
    int i = blockIdx.x * blockDim.x + threadIdx.x;
    out[i] = in[i] * coeffs[0];
}

// Kernel B
__global__ void useDifferentCoeff(float* out, const float* in) {
    int i = blockIdx.x * blockDim.x + threadIdx.x;
    out[i] = in[i] * coeffs[threadIdx.x % 256];
}

11 of 46 in GPU / CUDA