如何创建指向可与std :: nth_element和openMP一起使用的RcppEigen矩阵的指针?

时间:2018-08-25 09:16:41

标签: openmp eigen rcpp

我正在尝试在Rcpp中实现一个函数,该函数将矩阵作为输入,并根据用户为所述矩阵的行指定的方式计算和分位数。由于我想使用openMP,因此出于线程安全方面的考虑,我尝试使用RcppEigen进行操作。 看起来有点复杂的一个原因是,为了有效地计算分位数,我尝试模仿这种方法(finding quartiles,第一个答案),但允许用户输入。因此,基本上,我在第一步中创建了一个具有与分位数相对应的索引的向量。在第二步中,我尝试在for循环中访问相应的值。

这是我尝试的代码:

// // -*- mode: C++; c-indent-level: 4; c-basic-offset: 4; indent-tabs-mode: nil; -*-

// [[Rcpp::depends(RcppEigen)]]
#include <RcppEigen.h>
// [[Rcpp::plugins(openmp)]]

#ifdef _OPENMP
#include <omp.h>
#endif

// [[Rcpp::plugins(cpp11)]]
#include <random>

// [[Rcpp::export]]
SEXP summaryParC(const Eigen::MatrixXd x,
                 const Eigen::VectorXd quantiles,
                 int nrow, int ncol, const int ncores)
{
  const int no_quantiles = quantiles.size();
  Eigen::MatrixXd result(nrow, no_quantiles);

  // this part is just to give me a vector of indices I need later on in the foor loop
  //-----------------------------------------------
  Eigen::VectorXi indices(no_quantiles +1);
  indices[0] = -1;
  for (int k=0; k<no_quantiles; k++){
    if (quantiles[k] < 0.5){
      indices[k+1] = floor(quantiles[k] * (ncol-1));
    } else {
      indices[k+1] = ceil(quantiles[k] * (ncol-1));
    }
  }
  //-----------------------------------------------

#pragma omp parallel num_threads(ncores)
{
#pragma omp for
  for(int i = 0; i < nrow; i++){
    // I am trying to convert it into a vector so I can sort it
    Eigen::VectorXd v = (x.row(i));
    auto * ptr = v; // this fails
    // here I want to use the pointer to access the n-th element of the vector
    for(int q=0; q<no_quantiles; q++){ //quantiles
      std::nth_element(ptr + indices[q] + 1, ptr + indices[q+1], ptr + ncol);
      result(i,q) = *(ptr + indices[q+1]);
    }
  }
}
return Rcpp::wrap(result);
}

我想定义自己的指针的原因是Eigen :: VectorXd v没有像v.begin()那样的东西。如果没有openMP,我将简单地将x定义为NumericMatrix并将v定义为NumericVector,一切正常。使用openMP,我不能依赖于线程安全吗?

这适用于较小的数据集,但在较大的矩阵上使用时会崩溃:

// [[Rcpp::export]]
SEXP summaryC(NumericMatrix x,
                 NumericVector quantiles, 
                 int nrow, int ncol, const int ncores)
{
  const int no_quantiles = quantiles.size();
  NumericMatrix result(nrow, no_quantiles);
  int indices[no_quantiles +1];
  //-----------------------------------------------
  indices[0] = -1;
  for (int k=0; k<no_quantiles; k++){
    if (quantiles[k] < 0.5){
      indices[k+1] = floor(quantiles[k] * (ncol-1));
    } else {
      indices[k+1] = ceil(quantiles[k] * (ncol-1));
    }
  }
  //-----------------------------------------------
#pragma omp parallel num_threads(ncores)
{
#pragma omp for
  for(int i = 0; i < nrow; i++){
    // converting it into a vector so I can sort it
    NumericVector v = (x.row(i));
    for(int q=0; q<no_quantiles; q++){ //quantiles
      std::nth_element(v.begin() + indices[q] + 1, v.begin() + indices[q+1], v.end());
      result(i,q) = *(v.begin() + indices[q+1]);
    }
  }
}
  return Rcpp::wrap(result);
}

非常感谢您!

更新

我实现了Ralf Stubner的方法。据我所知,该指针工作正常。 (不幸的是,当我尝试运行R时,R仍会中止该会话。正如Dirk Eddelbuettel指出的那样,使用指针不能解决访问R内存的问题)。

// [[Rcpp::export]]
SEXP summaryParC(Eigen::MatrixXd x,
                 const Eigen::VectorXd quantiles,
                 int nrow, int ncol, const int ncores)
{
  const int no_quantiles = quantiles.size();
  Eigen::MatrixXd result(nrow, no_quantiles);
  Eigen::VectorXi indices(no_quantiles +1);
  indices[0] = -1;
  for (int k=0; k<no_quantiles; k++){
    if (quantiles[k] < 0.5){
      indices[k+1] = floor(quantiles[k] * (ncol-1));
    } else {
      indices[k+1] = ceil(quantiles[k] * (ncol-1));
    }
  }

#pragma omp parallel num_threads(ncores)
{
#pragma omp for
  for(int i = 0; i < nrow; i++){
    Eigen::VectorXd v = (x.row(i));
    double * B = v.data();
    double * E = B + nrow;

    for(int q=0; q<no_quantiles; q++){ //quantiles
      std::nth_element(B + indices[q] + 1, B + indices[q+1], E);
      result(i,q) = *(B + indices[q+1]);
    }
  }
}
return Rcpp::wrap(result);
}

第二次更新:这里是潜在问题的更清晰示例。我知道使用R结构在openMP中存在问题,但是该示例可能会导致人们更好地理解其根本原因。

// [[Rcpp::plugins(openmp)]]
// [[Rcpp::plugins(cpp11)]]
#include <Rcpp.h>
#ifdef _OPENMP
#include <omp.h>
#endif

using namespace Rcpp;

// [[Rcpp::export]]
SEXP summaryC(NumericMatrix x,
              int nrow, int ncol, const int ncores)
{
  NumericMatrix result(nrow, 5);
  int indices[6] = {-1, 0,  249,  500,  750, 999};

  //   #pragma omp parallel num_threads(ncores)
  {
    //     #pragma omp for
    for(int i = 0; i < nrow; i++){
      NumericVector v = (x.row(i));
      for(int q=0; q < 5; q++){
        std::nth_element(v.begin() + indices[q] + 1, v.begin() + indices[q+1], v.end());
        result(i,q) = *(v.begin() + indices[q+1]);
      }
    }
  }
  return Rcpp::wrap(result);
}





// [[Rcpp::export]]
SEXP summaryParC(NumericMatrix x,
                 int nrow, int ncol, const int ncores)
{
  NumericMatrix result(nrow, 5);
  int indices[6] = {-1, 0,  249,  500,  750, 999};

  #pragma omp parallel num_threads(ncores)
  {
    #pragma omp for schedule(dynamic)
      for(int i = 0; i < nrow; i++){
      {
        NumericVector v = (x.row(i));
        for(int q=0; q<5; q++){
          std::nth_element(v.begin() + indices[q] + 1, v.begin() + indices[q+1], v.end());
          result(i,q) = *(v.begin() + indices[q+1]);
        }
      }
      }
  }
return Rcpp::wrap(result);
}





// [[Rcpp::export]]
SEXP summaryParCorder(NumericMatrix x,
                 int nrow, int ncol, const int ncores)
{
  NumericMatrix result(nrow, 5);
  int indices[6] = {-1, 0,  249,  500,  750, 999};

  #pragma omp parallel num_threads(ncores)
  {
    #pragma omp for ordered schedule(dynamic)
    for(int i = 0; i < nrow; i++){
      #pragma omp ordered
      {
        NumericVector v = (x.row(i));
        for(int q=0; q<5; q++){
          std::nth_element(v.begin() + indices[q] + 1, v.begin() + indices[q+1], v.end());
          result(i,q) = *(v.begin() + indices[q+1]);
        }
      }
    }
  }
return Rcpp::wrap(result);
}




***** R - code *****
#this works, but summaryParCorder is much slower. 
mbm <- microbenchmark::microbenchmark(
  summaryC(x = matrix(as.numeric(1:1000000), ncol = 1000), 
           nrow = 1000, ncol = 1000, ncores = 4),

  summaryParCorder(x = matrix(as.numeric(1:1000000), ncol = 1000), 
              nrow = 1000, ncol = 1000, ncores = 4),
  times = 20
)
mbm

# this breaks:
summaryParC(x = matrix(as.numeric(1:1000000), ncol = 1000), 
                 nrow = 1000, ncol = 1000, ncores = 4)

1 个答案:

答案 0 :(得分:1)

我尚未检查与OpenMP的兼容性,但是如果所考虑的向量不是Eigen::VectorXd::data(),则const会为您提供所需的指针:

// [[Rcpp::depends(RcppEigen)]]
#include <RcppEigen.h>

// [[Rcpp::export]]
Eigen::VectorXd quantiles(Eigen::VectorXd x, const Eigen::VectorXi& indices) {
  Eigen::VectorXd result(indices.size());

  std::nth_element(x.data(), x.data() + indices[0], x.data() + x.size());
  result(0) = x[indices[0]];

  for (int i = 1; i < indices.size(); ++i) {
    std::nth_element(x.data() + indices[i - 1] + 1,
                     x.data() + indices[i],
                     x.data() + x.size());
    result(i) = x[indices[i]];
  }
  return result;
}

/*** R
set.seed(42)
x <- runif(12)
i <- sort(sample(seq_len(12), 3)) - 1
quantiles(x, i)
*/

这里是包括OpenMP在内的完整解决方案:

// [[Rcpp::plugins(openmp)]]
// [[Rcpp::plugins(cpp11)]]
// [[Rcpp::depends(RcppEigen)]]
#include <RcppEigen.h>

using namespace Rcpp;

// [[Rcpp::export]]
NumericMatrix summaryC(NumericMatrix x, int nrow, int ncores)
{
  NumericMatrix result(nrow, 5);
  int indices[6] = {-1, 0,  249,  500,  750, 999};

  for (int i = 0; i < nrow; i++) {
    NumericVector v = (x.row(i));
    for (int q = 0; q < 5; ++q) {
      std::nth_element(v.begin() + indices[q] + 1, v.begin() + indices[q+1], v.end());
      result(i,q) = *(v.begin() + indices[q+1]);
    }
  }
  return result;
}

// [[Rcpp::export]]
Eigen::MatrixXd summaryParC(Eigen::MatrixXd x,int nrow, int ncores) {
  Eigen::MatrixXd result(nrow, 5);
  int indices[6] = {-1, 0,  249,  500,  750, 999};

  #pragma omp parallel num_threads(ncores)
  {
    #pragma omp for schedule(dynamic)
      for (int i = 0; i < nrow; i++) {
        Eigen::VectorXd v = x.row(i);
        for (int q = 0; q < 5; ++q) {
          std::nth_element(v.data() + indices[q] + 1,
               v.data() + indices[q+1],
               v.data() + v.size());
          result(i,q) = v[indices[q+1]];
        }
      }
  }
  return result;
}

/*** R 
x <- matrix(as.numeric(1:1000000), ncol = 1000)
microbenchmark::microbenchmark(
   summaryC = summaryC(x = x, nrow = 1000, ncores = 4),
  summaryParC = summaryParC(x = x, nrow = 1000, ncores = 4),
  times = 100)
*/

我从未见过此并行版本崩溃。在我的双核计算机上,它比串行代码快44%。