TaskManager* createTaskManagerImpl(struct Context* context, int numCPU, int numGPU, int numIO) {
    struct TaskManager* taskManager = new TaskManager();
    taskManager->spawnTasks = C_spawnTasksTaskManagerImpl;
    taskManager->spawn = C_spawnTaskManagerImpl;
    taskManager->shutdown  = C_shutdownTaskManagerImpl;
    taskManager->incrementTaskCount = C_incrementTaskCountTaskManagerImpl;
    taskManager->decrementTaskCount = C_decrementTaskCountTaskManagerImpl;
    taskManager->setWaitTask = C_setWaitTaskTaskManagerImpl;
    struct TaskManagerImpl* taskManagerImpl = new TaskManagerImpl();
    // 0...numIO-1 IOProcessor
    // numIO...numIO+numGPU-1 GPUProcessor
    // numIO+numGPU...numIO+numGPU+numCPU-1 CPUProcessor
    taskManagerImpl->io = 0;
    taskManagerImpl->gpu = numIO;
    taskManagerImpl->cpu = numIO+numGPU;
    taskManagerImpl->maxCPU = numIO+numGPU+numCPU;
    taskManagerImpl->numWorker = taskManagerImpl->maxCPU;
    taskManagerImpl->sendGPUWorkerIndex = taskManagerImpl->gpu;
    taskManagerImpl->sendCPUWorkerIndex = taskManagerImpl->cpu;
    taskManagerImpl->taskCount = 0;
    taskManagerImpl->loopCounter = 0;
    createWorkers(context, taskManagerImpl);
    taskManager->taskManager = (union Data*)taskManagerImpl;
    return taskManager;
}

void createWorkers(struct Context* context, TaskManagerImpl* taskManager) {
    int i = 0;
    taskManager->workers = (Worker**)ALLOCATE_PTR_ARRAY(context, Worker, taskManager->maxCPU);
    for (;i<taskManager->gpu;i++) {
        Queue* queue = createSynchronizedQueue(context);
        taskManager->workers[i] = (Worker*)createCPUWorker(context, i, queue);
    }
    for (;i<taskManager->cpu;i++) {
        Queue* queue = createSynchronizedQueue(context);
#ifdef USE_CUDAWorker
        taskManager->workers[i] = (Worker*)createCUDAWorker(context, i, queue,0);
#else
        taskManager->workers[i] = (Worker*)createCPUWorker(context, i, queue);
#endif
    }
    for (;i<taskManager->maxCPU;i++) {
        Queue* queue = createSynchronizedQueue(context);
        taskManager->workers[i] = (Worker*)createCPUWorker(context, i, queue);
    }
}