------------
--------++--
--------++--
------------
------------

// We break the threads up into four flights:
//
//     0: even X and even Y
//     1:  odd X and even Y
//     2: even X and  odd Y
//     3:  odd X and  odd Y
const int flight = ( threadIdx.x % 2 + ( ( threadIdx.y % 2 ) << 1 ) );

for (int idx = 0; idx < flight; ++idx) {
  __syncthreads();
 }

doWork( pixel[ threadIdx.x +     threadIdx.y * blockDim.x ],
        pixel[ threadIdx.x + 1 + threadIdx.y * blockDim.x ],
        pixel[ threadIdx.x + 1 + (threadIdx.y + 1) * blockDim.x ],
        pixel[ threadIdx.x +     (threadIdx.y + 1) * blockDim.x ]);

for (int idx = 3; idx > flight; --idx) {
  __syncthreads();
 }

doWork(...);
__syncthreads("one");
__syncthreads("two");
__syncthreads("three");

__syncthreads("one");
doWork(...);
__syncthreads("two");
__syncthreads("three");

__syncthreads("one");
__syncthreads("two");
doWork(...);
__syncthreads("three");

__syncthreads("one");
__syncthreads("two");
__syncthreads("three");
doWork(...);