在 CUDA 中使用推力::排序时出现分段错误 [英] Segmentation error when using thrust::sort in CUDA

问题描述

我试图通过将比较函数作为参数传递给推力排序，根据其类型对类对象数组进行排序.

I am trying to sort an array of class objects based on its type by passing a comparison function as the parameter to the thrust sort.

类定义:

class TetraCutInfo
{

        public:
        int tetraid;
        unsigned int ncutEdges;
        unsigned int ncutNodes;
        unsigned int type_cut;
        __host__ __device__ TetraCutInfo();
};

排序:

   thrust::sort(cutInfoptr,cutInfoptr+n,cmp());

cutInfoptr 是 TetraCutInfo 类型的指针，具有使用 cudaMalloc 分配的设备内存地址.

cutInfoptr is a pointer of type TetraCutInfo having the address of the device memory allocated using cudaMalloc.

比较功能

struct cmp
{
  __host__ __device__
  bool operator()(const TetraCutInfo x, TetraCutInfo y)
  {
        return (x.type_cut < y.type_cut);
  }
};

在运行它时，我遇到了分段错误，但是我能够在另一个内核中迭代 cutInfoptr.

On running this I am getting Segmentation fault, however I am able to iterate through cutInfoptr in another kernel.

PS: 我参考了链接 https://code.google.com/p/thrust/source/browse/examples/sort.cu

推荐答案

cutInfoptr 是 TetraCutInfo 类型的指针，具有使用 cudaMalloc 分配的设备内存地址.

cutInfoptr is a pointer of type TetraCutInfo having the address of the device memory allocated using cudaMalloc.

虽然您没有展示完整的代码，但根据您所做的上述声明，事情可能无法正常工作，并且我预计会出现段错误，因为该指针被取消引用.

Although you haven't shown a complete code, based on the above statement you made, things probably won't work, and I would expect a seg fault as that pointer gets dereferenced.

请注意 thrust 快速入门指南中提供的信息:

Note the information given in the thrust quick start guide:

您可能想知道当原始"指针用作推力函数的参数.与 STL 一样，Thrust 允许这种用法，它将分发算法的主机路径.如果有问题的指针实际上是指向设备内存的指针，那么您需要在调用该函数之前用推力::device_ptr 包装它.

You may wonder what happens when a "raw" pointer is used as an argument to a Thrust function. Like the STL, Thrust permits this usage and it will dispatch the host path of the algorithm. If the pointer in question is in fact a pointer to device memory then you'll need to wrap it with thrust::device_ptr before calling the function.

您引用的 cutInfoptr，如果由 cudaMalloc 创建，则是原始指针"；(这也恰好是一个设备指针).当您将它传递给推力时，推力会看到它是一个原始指针，并分派主机路径".当您传递的(设备)指针在主机路径中的主机代码中被取消引用时，您会遇到段错误.

The cutInfoptr you referenced, if being created by cudaMalloc, is a "raw pointer" (which also happens to be a device pointer). When you pass it to thrust, thrust sees that it is a raw pointer, and dispatches the "host path". When the (device) pointer you pass is dereferenced in host code in the host path, you get a seg fault.

一种解决方案是将其包裹在一个推力::device_ptr指针中，这里摘录快速入门指南示例:

One solution is to wrap it in a thrust::device_ptr pointer, excerpting the quick start guide example here:

size_t N = 10;

// raw pointer to device memory
int * raw_ptr;
cudaMalloc((void **) &raw_ptr, N * sizeof(int));

// wrap raw pointer with a device_ptr 
thrust::device_ptr<int> dev_ptr(raw_ptr);

// use device_ptr in thrust algorithms
thrust::fill(dev_ptr, dev_ptr + N, (int) 0);

另一种可能的解决方案是使用适当的执行策略进行分派，例如<代码>推力::设备.

Another possible solution is to dispatch with an appropriate execution policy, such as thrust::device.

这篇关于在 CUDA 中使用推力::排序时出现分段错误的文章就介绍到这了，希望我们推荐的答案对大家有所帮助，也希望大家多多支持IT屋！