c++ - 如何在 C++ 或 OpenCV SubMatrix 中的 2D 子 vector 上使用 OpenACC？

Question

我有以下代码

int main(int argc, char** argv )
{
    std::cout<<"running Lenna..\n";
    cv::Mat mat = imread("lena.bmp", cv::IMREAD_GRAYSCALE );

    //convert to vec
    std::vector<double> BWvec;
    BWvec.assign((double*)mat.data, (double*)mat.data + mat.total());
    std::vector < std::vector<double>> vec2D;
    for (int i = 0; i < mat.rows; i++) {
        auto first = BWvec.begin() + (mat.rows * i);
        auto last = BWvec.begin() + (mat.rows * i) + mat.rows;
        std::vector<double> vec0(first, last);
        vec2D.push_back(vec0);
    }

    //#pragma acc parallel loop
    for (int i = 0; i <= 5; i++) {
        for (int j = 0; j <= 5; j++) {
            mat(cv::Rect(i,j, (4 - 0), (4 - 0)));

            //sub-vector[5:10][25:100]:
            std::vector<std::vector<double>> sub_vector;
            sub_vector.reserve(5);
            for (std::size_t k = 5; k < 10; ++k) {
                sub_vector.emplace_back(vec2D[k+i].begin() + 25, vec2D[k+i].begin() + 100);
            }
        }
    }

    return 0;
}

当我输入 pgc++ -fast -ta=nvidia:cuda9.2,managed -Minfo=accel -o lenna lenna.cpp -std=c++11 pkg-config --cflags --libs opencv -lgomp && ./lenna ，它可以正常工作，但是当我取消注释 #pragma acc parallel loop ，我得到错误

procedures called in a compute region must have acc routine information
accelerator region ignored, accelerator restriction .. no acc routine information

如果我注释掉 mat(cv::Rect(i,j,(4-0),(4-0)))，我也会收到此错误并在 sub-vector[5:10][25:100] 之后留下部分，或者如果我取消注释 mat(cv::Rect(i,j,(4-0),(4-0)))并评论 sub-vector[5:10][25:100] 之后的部分
我怎样才能解决这个问题？
编辑
为了使这更简单，我提供了 2 个单独的代码，以及它们给出的错误: lenna1.cpp :

#include <stdio.h>
#include <cmath>
#include <omp.h>
#include <opencv2/opencv.hpp>

using namespace std;
using namespace cv;

//pgc++ -fast -ta=nvidia:cuda9.2,managed -Minfo=accel -o lenna lenna.cpp -std=c++11 `pkg-config --cflags --libs opencv` -lgomp && ./lenna

int main(int argc, char** argv )
{
    std::cout<<"running Lenna..\n";
    cv::Mat mat = imread("lena.bmp", cv::IMREAD_GRAYSCALE );

    #pragma acc parallel loop
    for (int i = 0; i <= 5; i++) {
        for (int j = 0; j <= 5; j++) {
            mat(cv::Rect(i, j, (4 - 0), (4 - 0)));
        }
    }
    return 0;
}

来自 lenna1.cpp 的错误:

pgc++ -fast -ta=nvidia:cuda9.2,managed -Minfo=accel -o lenna1 lenna1.cpp -std=c++11 `pkg-config --cflags --libs opencv` -lgomp && ./lenna1
lenna1.cpp:
"lenna1.cpp", line 23: warning: last line of file ends without a newline
  }
   ^

PGCC-S-0155-Procedures called in a compute region must have acc routine information: cv::Mat::Mat(const cv::Mat&, const cv::Rect_<int> &) (lenna1.cpp: 379)
PGCC-S-0155-Accelerator region ignored; see -Minfo messages  (lenna1.cpp: 14)
main:
     14, Accelerator region ignored
         379, Accelerator restriction: call to 'cv::Mat::Mat(const cv::Mat&, const cv::Rect_<int> &)' with no acc routine information
PGCC/x86-64 Linux 19.10-0: compilation completed with severe errors

lenna2.cpp :

#include <stdio.h>
#include <cmath>
#include <omp.h>
#include <opencv2/opencv.hpp>

using namespace std;
using namespace cv;

//pgc++ -fast -ta=nvidia:cuda9.2,managed -Minfo=accel -o lenna lenna.cpp -std=c++11 `pkg-config --cflags --libs opencv` -lgomp && ./lenna

int main(int argc, char** argv )
{
    std::cout<<"running Lenna..\n";
    cv::Mat mat = imread("lena.bmp", cv::IMREAD_GRAYSCALE );

    //convert to vec
    std::vector<double> BWvec;
    BWvec.assign((double*)mat.data, (double*)mat.data + mat.total());
    std::vector < std::vector<double>> vec2D;
    for (int i = 0; i < mat.rows; i++) {
        auto first = BWvec.begin() + (mat.rows * i);
        auto last = BWvec.begin() + (mat.rows * i) + mat.rows;
        std::vector<double> vec0(first, last);
        vec2D.push_back(vec0);
    }

    #pragma acc parallel loop
    for (int i = 0; i <= 5; i++) {
        for (int j = 0; j <= 5; j++) {
            //sub-vector[5:10][25:100]:
            std::vector<std::vector<double>> sub_vector;
            sub_vector.reserve(5);
            for (std::size_t i = 5; i < 10; ++i) {
                sub_vector.emplace_back(vec2D[i].begin() + 25, vec2D[i].begin() + 100);
            }
        }
    }

    return 0;
}

来自 lenna2.cpp 的错误:

pgc++ -fast -ta=nvidia:cuda9.2,managed -Minfo=accel -o lenna2 lenna2.cpp -std=c++11 `pkg-config --cflags --libs opencv` -lgomp && ./lenna2
lenna2.cpp:
"lenna2.cpp", line 40: warning: last line of file ends without a newline
  }
   ^

operator new (unsigned long, void *):
      4, include "opencv.hpp"
          47, include "core.hpp"
               56, include "algorithm"
                    10, include "algorithm"
                         62, include "stl_algo.h"
                              62, include "stl_tempbuf.h"
                                   60, include "stl_construct.h"
                                        59, include "new"
                                            130, Generating implicit acc routine seq
                                                 Generating acc routine seq
                                                 Generating Tesla code
operator delete (void *, void *):
      4, include "opencv.hpp"
          47, include "core.hpp"
               56, include "algorithm"
                    10, include "algorithm"
                         62, include "stl_algo.h"
                              62, include "stl_tempbuf.h"
                                   60, include "stl_construct.h"
                                        59, include "new"
                                            135, Generating implicit acc routine seq
                                                 Generating acc routine seq
                                                 Generating Tesla code
PGCC-S-0155-Procedures called in a compute region must have acc routine information: std::__throw_length_error(const char *) (lenna2.cpp: 69)
PGCC-S-0155-Accelerator region ignored; see -Minfo messages  (lenna2.cpp: 25)
main:
     25, Accelerator region ignored
          69, Accelerator restriction: call to 'std::__throw_length_error(const char *)' with no acc routine information
PGCC/x86-64 Linux 19.10-0: compilation completed with severe errors

Answer 1

为了从设备调用例程和方法，需要有这些例程的设备版本。在已知调用例程的定义的情况下(例如使用模板)，编译器将尝试隐式生成设备例程。否则，程序员有责任用 OpenACC“例程”指令装饰被调用的例程。
由于您提供的信息不完整，因此很难确切知道如何修复您的代码。错误消息说缺少哪些例程？你能提供一个完整的复制例子吗？
更新后编辑。

call to 'cv::Mat::Mat(const cv::Mat&, const cv::Rect_ &)' with noacc routine information

看起来“Mat”类型的构造函数没有设备可调用版本。虽然我不熟悉 OpenCV 的结构，但我假设这不是模板化的，也不是头文件中包含的构造函数的定义，因此编译器可以隐式创建它。您需要将例程指令添加到您希望从设备代码调用的 OpenCV 部分，或者如果有 CUDA 设备例程，您可以通过使用带有绑定(bind)子句的 OpenACC 例程指令来调用它们。

69, Accelerator restriction: call to 'std::__throw_length_error(constchar *)' with no acc routine information

异常处理不适用于设备代码，因为它需要在主机上捕获，而目前没有办法支持这一点。
在某些情况下，您可以通过标志“--no_exceptions”禁用异常来解决此问题，但在这种情况下，如果禁用异常，OpenCV 会提示。因此，最好避免在此处在设备上使用 vector 。