gistfile1.c

    uint i = gl_LocalInvocationID.x; // 0 <= i < 64
    int coeff = 0;
    uint nonzero_lo = ac_nonzero_lo[block_idx];
    uint nonzero_hi = ac_nonzero_hi[block_idx];
    uint lane_mask = 1 << (i & 31);
    if (i < 32)
    {
        if ((nonzero_lo & lane_mask) != 0)
        {
            uint fetch_offs = ptr + bitCount(nonzero_lo & (lane_mask - 1));
            coeff = get_int16(fetch_offs);
        }
    }
    else
    {
        if ((nonzero_hi & lane_mask) != 0)
        {
            uint fetch_offs = ptr + bitCount(nonzero_lo) + bitCount(nonzero_hi & (lane_mask - 1));
            coeff = get_int16(fetch_offs);
        }
    }