如何使用 OpenACC 将 C++ 对象数组正确复制到 GPU？

如何解决如何使用 OpenACC 将 C++ 对象数组正确复制到 GPU？

我正在尝试使用 OpenACC 将一组 C++ 对象从主机复制到 GPU 设备。我试过在网上查找，但只找到了关于如何完成此操作的非常模糊的文档，因此不确定是否有人知道正确的方法？我还将 nvhpc pgi 编译器与 NVIDIA gpu 一起使用，因此我可以访问托管内存，想要使用手动传输和/或托管的示例。

假设我有一个简单的作业类（关于学校作业）：

class Assignment{
Public:
   string name; //name of assignment
   double *grades; // list of grades where n is the number of students

   Assignment(string name_,double *grades_) {
      name = name_;
      grades = grades_;
   }
}


int main(){
   Assignment assignments[2]; // array of 2 assignments

   double grades1[3] = {90.0,95.0,75.0};
   assignments[0] = Assignment("Assignment1",grades1);

   double grades2[3] = {50.0,65.0,55.0};
   assignments[1] = Assignment("Assignment2",grades2);

   //Transfer to GPU below...
   #pragma ...
}

如何使用 OpenACC 手动内存管理将这个赋值 C++ 对象数组传输到 GPU？和/或还使用托管内存？

解决方法

统一内存（即托管）可用于分配的内存。因此，如果您更改代码以动态分配数组并使用标志“-gpu=managed”进行编译，则数组的数据移动将由驱动程序处理。

OpenACC 数据区域将执行浅拷贝。因此，对于具有动态数据成员的聚合数据类型，您需要进行如下所示的手动深层复制：

% cat grades.cpp
#include <iostream>
#include <string>

using namespace std;

class Assignment{
public:
   std::string name; //name of assignment
   double *grades; // list of grades where n is the number of students

   Assignment() {};
   Assignment(string name_,double *grades_) {
      name = name_;
      grades = grades_;
   }
};


int main(){
   Assignment assignments[2]; // array of 2 assignments

   double grades1[3] = {90.0,95.0,75.0};
   assignments[0] = Assignment("Assignment1",grades1);

   double grades2[3] = {50.0,65.0,55.0};
   assignments[1] = Assignment("Assignment2",grades2);

   //Transfer to GPU below...
   #pragma acc enter data copyin(assignments[:2])
   for (int i=0; i < 2; ++i) {
        #pragma acc enter data copyin(assignments[i].grades[:3])
   }
   // changes the grades on the device
   #pragma acc parallel loop present(assignments)
   for (int i=0; i < 2; ++i) {
      for (int j=0; j < 3; ++j) {
         assignments[i].grades[j] += 5.0;
      }
   }

   for (int i=0; i < 2; ++i) {
        #pragma acc update self(assignments[i].grades[:3])
        std::cout << i << " : " << assignments[i].grades[0]
                  << "," <<  assignments[i].grades[1]
                  << "," <<  assignments[i].grades[2]
                  << std::endl;
   }


   for (int i=0; i < 2; ++i) {
        #pragma acc exit data delete(assignments[i].grades)
   }
   #pragma acc exit data delete(assignments)

}
% nvc++ grades.cpp -w -Minfo=accel -acc; a.out
main:
     26,Generating enter data copyin(assignments[:])
     32,Generating enter data copyin(assignments.grades[:3])
         Generating present(assignments[:])
         Generating Tesla code
         35,#pragma acc loop gang /* blockIdx.x */
         36,#pragma acc loop seq
     36,Loop is parallelizable
     43,Generating update self(assignments.grades[:3])
     52,Generating exit data delete(assignments.grades)
     55,Generating exit data delete(assignments[:])
0 : 95,100,80
1 : 55,70,60

我个人更喜欢将数据管理封装到类本身中。这里有点棘手，因为您有一个类数组，其中每个元素都是匿名对象的副本。虽然它可能有用也可能没用，但您可以查看一个简单的向量，如类“accList”，我将其作为我在Parallel Programing with OpenACC一书中关于数据管理的章节的示例>.请参阅：https://github.com/rmfarber/ParallelProgrammingWithOpenACC/tree/master/Chapter05