GPU上的CUDA内核调度程序

Question

I'm writing a CUDA kernel scheduler. The scheduler gets a vector of Task pointers and brings them to execution. The pointers point to KernelTask objects of different type parameters to support kernels with arbitrary parameters.

There's a CPU version of the Scheduler and a GPU version. The CPU version works just fine. It calls the virtual function Task::start to execute a Kernel. The GPU version has three problems:

1. Virtual functions are not allowed in CUDA. How can I avoid them without down casting?
2. std::get is a host function. Is there a way to implement std::get myself for the GPU?
3. (Low priority) Because KernelTask objects are of diffrent size I copy all of them seperatly with copyToGPU() . Is there a way for batch copying?

Here is the code:

// see http://stackoverflow.com/questions/7858817/unpacking-a-tuple-to-call-a-matching-function-pointer
template<int ...>
struct seq { };

template<int N, int ...S>
struct gens : gens<N-1, N-1, S...> { };

template<int ...S>
struct gens<0, S...> {
  typedef seq<S...> type;
};

class Task {
private:
    bool visited;
    bool reached;
protected:
    std::vector<std::shared_ptr<Task>> dependsOn;
    Task();
public:
    Task **d_dependsOn = NULL;
    int d_dependsOnSize;
    Task *d_self = NULL;

    int streamId;
    int id;
    cudaStream_t stream;

    virtual void copyToGPU() = 0;
    virtual void start() = 0;
    virtual void d_start() = 0;
    virtual ~Task() {}
    void init();
    void addDependency(std::shared_ptr<Task> t);
    cudaStream_t dfs();
};

template<typename... Args>
class KernelTask : public Task {
private:
    std::tuple<Args...> params;
    dim3 threads;
    dim3 blocks;
    void (*kfp)(Args...);

    template<int ...S>
    void callFunc(seq<S...>) {
        // inserting task into stream
        this->kfp<<<this->blocks, this->threads, 0, this->stream>>>(std::get<S>(params) ...);
        checkCudaErrors(cudaGetLastError());

        if (DEBUG) printf("Task %d: Inserting Task in Stream.\n", this->id);
    }

    template<int ...S>
    __device__ void d_callFunc(seq<S...>) {
        // inserting task into stream
        this->kfp<<<this->blocks, this->threads, 0, this->stream>>>(std::get<S>(params) ...);

        if (DEBUG) printf("Task %d: Inserting Task in Stream.\n", this->id);
    }

    KernelTask(int id, void (*kfp)(Args...), std::tuple<Args...> params, dim3 threads, dim3 blocks);

public:
    ~KernelTask();
    void copyToGPU();

    void start() override {
        callFunc(typename gens<sizeof...(Args)>::type());
    }

    __device__ void d_start() override {
        d_callFunc(typename gens<sizeof...(Args)>::type());
    }

    static std::shared_ptr<KernelTask<Args...>> create(int id, void (*kfp)(Args...), std::tuple<Args...> params, dim3 threads, dim3 blocks);
};

class Scheduler {
private:
    std::vector<std::shared_ptr<Task>> tasks;
public:
    Scheduler(std::vector<std::shared_ptr<Task>> &tasks) {
        this->tasks = tasks;
    }

    void runCPUScheduler();
    void runGPUScheduler();
};

---

EDIT:

(1) Virtual Functions in CUDA: I get a Warp Illegal Address exception in scheduler in the following example:

struct Base {
    __host__ __device__ virtual void start() = 0;
    virtual ~Base() {}
};

struct Derived : Base {
    __host__ __device__ void start() override {
        printf("In start\n");
    }
};

__global__ void scheduler(Base *c) {
    c->start();
}

int main(int argc, char **argv) {
    Base *c = new Derived();
    Base *d_c;
    checkCudaErrors(cudaMalloc(&d_c, sizeof(Derived)));
    checkCudaErrors(cudaMemcpy(d_c, c, sizeof(Derived), cudaMemcpyHostToDevice));

    c->start();
    scheduler<<<1,1>>>(d_c);

    checkCudaErrors(cudaFree(d_c));

    return 0;
}

(2) thrust::tuple works fine.

(3) I'm open to suggestions.

(4) How do I pass a kernel function pointer to a kernel? I get a Warp Misaligned Address exception in the following example:

__global__ void baz(int a, int b) {
    printf("%d + %d = %d\n", a, b, a+b);
}

void schedulerHost(void (*kfp)(int, int)) {
    kfp<<<1,1>>>(1,2);
}

__global__ void schedulerDevice(void (*kfp)(int, int)) {
    kfp<<<1,1>>>(1,2);
}

int main(int argc, char **argv) {
    schedulerHost(&baz);
    schedulerDevice<<<1,1>>>(&baz);
    return 0;
}

Answer 1

"Virtual functions are not allowed in CUDA. How can I avoid them without down casting?"

You can have both virtual __host__ and __device__ functions: http://docs.nvidia.com/cuda/cuda-c-programming-guide/#virtual-functions

However:

It is not allowed to pass as an argument to a __global__ function an object of a class with virtual functions.

---

"std::get is a host function. Is there a way to implement std::get myself for the GPU? "

I'd suggest using thrust::tuple instead which has both a __host__ and a __device__ implementation: http://thrust.github.io/doc/group__tuple.html

---

Regarding function pointers:

The address of a __global__ function taken in host code cannot be used in device code (eg to launch the kernel).

http://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#function-pointers