How to implement device side CUDA virtual functions?

淺唱寂寞╮ 提交于 2021-02-02 08:13:54

问题


I see that CUDA doesn't allow for classes with virtual functions to be passed into kernel functions. Are there any work-arounds to this limitation?

I would really like to be able to use polymorphism within a kernel function.

Thanks!


回答1:


The most important part of Robert Crovella's comment is:

The objects simply need to be created on the device.

So keeping that in mind, I was dealing with situation where I had an abstract class Function and then some implementations of it encapsulating different function and its evaluation. This is the simplified version of my code how I achieved polymorphism in my situation, but I am not saying it cannot be done better... It will hopefully help you to get the idea:

class Function
{
public:
    __device__ Function() {}
    __device__ virtual ~Function() {}
    __device__ virtual void Evaluate(const real* __restrict__ positions, real* fitnesses, const SIZE_TYPE particlesCount) const = 0;
};

class FunctionRsj : public Function
{
private:
    SIZE_TYPE m_DimensionsCount;
    SIZE_TYPE m_PointsCount;
    real* m_Y;
    real* m_X;
public:
    __device__ FunctionRsj(const SIZE_TYPE dimensionsCount, const SIZE_TYPE pointsCount, real* configFileData)
        : m_DimensionsCount(dimensionsCount),
            m_PointsCount(pointsCount),
            m_Y(configFileData),
            m_X(configFileData + pointsCount) {}

    __device__ ~FunctionRsj()
    {
        // m_Y points to the beginning of the config
        // file data, use it for destruction as this 
        // object took ownership of configFilDeata.
        delete[] m_Y;
    }

    __device__ void Evaluate(const real* __restrict__ positions, real* fitnesses, const SIZE_TYPE particlesCount) const
    {
        // Implement evaluation of FunctionRsj here.
    }
};

__global__ void evaluate_fitnesses(
    const real* __restrict__ positions,
    real* fitnesses,
    Function const* const* __restrict__ function,
    const SIZE_TYPE particlesCount)
{
    // This whole kernel is just a proxy as kernels
    // cannot be member functions.
    (*function)->Evaluate(positions, fitnesses, particlesCount);
}

__global__ void create_function(
    Function** function,
    SIZE_TYPE dimensionsCount,
    SIZE_TYPE pointsCount,
    real* configFileData)
{
    // It is necessary to create object representing a function
    // directly in global memory of the GPU device for virtual
    // functions to work correctly, i.e. virtual function table
    // HAS to be on GPU as well.
    if (threadIdx.x == 0 && blockIdx.x == 0)
    {
        (*function) = new FunctionRsj(dimensionsCount, pointsCount, configFileData);
    }
}

__global__ void delete_function(Function** function)
{
    delete *function;
}

int main()
{
    // Lets just assume d_FunctionConfigData, d_Positions,
    // d_Fitnesses are arrays allocated on GPU already ...

    // Create function.
    Function** d_Function;
    cudaMalloc(&d_Function, sizeof(Function**));
    create_function<<<1, 1>>>(d_Function, 10, 10, d_FunctionConfigData);

    // Evaluate using proxy kernel.
    evaluate_fitnesses<<<
        m_Configuration.GetEvaluationGridSize(),
        m_Configuration.GetEvaluationBlockSize(),
        m_Configuration.GetEvaluationSharedMemorySize()>>>(
        d_Positions,
        d_Fitnesses,
        d_Function,
        m_Configuration.GetParticlesCount());

    // Delete function object on GPU.
    delete_function<<<1, 1>>>(d_Function);
}


来源:https://stackoverflow.com/questions/26812913/how-to-implement-device-side-cuda-virtual-functions

易学教程内所有资源均来自网络或用户发布的内容,如有违反法律规定的内容欢迎反馈
该文章没有解决你所遇到的问题?点击提问,说说你的问题,让更多的人一起探讨吧!