为了设置上下文，我有一个正在运行的C++代码（一个图形算法），我试图在GPU和多核上使用OpenACC /nvc++来加速它。我使用vector或vector来存储邻接列表。由于向量会影响性能，我决定使用CSR数组来重写--这在GPU程序中最常见。目标是随机化每个顶点的邻居。
创建CSR阵列后，我使用std::shuffle来实现相同的操作。

int* offset;
  int* value;
  
  convert_VecOfVec_to_CSR(a, offset, value); 
  shuffleIt(offset, value, n);
  
  // DEFINITION
void shuffleIt(int *&offset, int* &value, int n){
  #pragma acc parallel loop
  for (int i = 0; i < n; ++i) {
    std::shuffle(value+offset[i], value+offset[i+1] ,std::default_random_engine(rand()));
  }
}

当使用acc=multicore时，代码编译和运行正常。但是我得到了下面的gpu错误。这是否意味着我不能在我的OpenACC的gpu代码中使用std::shuffle或random函数？我记得math.h函数工作正常。请提供建议。谢谢。

$ nvc++ -acc -gpu=managed -Minfo=all shuffle-2D-csr2.cpp -o shuffle-2D-csr2.out && ./shuffle-2D-csr2.out
NVC++-S-1061-Procedures called in a compute region must have acc routine information - std::linear_congruential_engine<unsigned long, (unsigned long)16807, (unsigned long)0, (unsigned long)2147483647>::linear_... (shuffle-2D-csr2.cpp: 66)
shuffleIt(int *&, int *&, int):
     66, Accelerator restriction: call to 'std::linear_congruential_engine<unsigned long, (unsigned long)16807, (unsigned long)0, (unsigned long)2147483647>::linear_congruential_engine(unsigned long)' with no acc routine information
std::linear_congruential_engine<unsigned long, (unsigned long)16807, (unsigned long)0, (unsigned long)2147483647>::operator ()():
      7, include "random"
          49, include "random.h"
              348, Generating implicit acc routine seq
                   Generating acc routine seq
                   Generating NVIDIA GPU code
...
NVC++/x86-64 Linux 22.7-0: compilation completed with severe errors

我还没有添加任何data子句，为了首先使功能正确。
Full Gist code
我正在学习OpenACC。