shuffle.cl

__kernel void shuffle_data(
    const __global float* data, // array to be shuffled
    const __global uint2* swaps  // array of swaps to perform
) {
    const int gid = get_global_id(0);
    const int field = get_global_id(1);
    const int num_fields = get_global_size(1);
    // swap row r1 with r2 from swaps[gid]
    uint2 swap = swaps[gid];
    const int r1 = num_fields* swap.x + field;
    const int r2 = num_fields* swap.y + field;
    float tmp = data[r2];
    data[r2] = data[r1];
    data[r1] = tmp;
}