c++ - 串行工作时Cuda版本不工作

Question

我的以下极简 Cuda 代码返回了不正确的结果（所有多边形最后都有 0 个顶点），而在 C++ 中串行运行的相同代码运行良好。问题是令人尴尬的并行：没有通信，没有同步线程等，而且 Cuda 内存分配是成功的。甚至我存储输入数组内容以进行调试的虚拟变量对于 Cuda 版本也是 0。由于我的数组足够大，因此没有越界访问。用 Cuda 中的循环替换 memcpy 不会改变任何东西。
我真的不明白会发生什么......任何想法？谢谢！

库达代码：

    #include <stdio.h>
    #include <iostream>
    #include <stdlib.h>
    #include <cuda.h>

    class Point2D {
     public:
     __device__ Point2D(double xx=0, double yy=0):x(xx),y(yy){};
     double x, y;
    };

    __device__ double dot(const Point2D &A, const Point2D &B) {
     return A.x*B.x + A.y*B.y;
    }
    __device__ Point2D operator*(double a, const Point2D &P) {
     return Point2D(a*P.x, a*P.y);
    }
    __device__ Point2D operator+(Point2D A, const Point2D &B) {
     return Point2D(A.x + B.x, A.y + B.y);
    }
    __device__ Point2D operator-(Point2D A, const Point2D &B) {
     return Point2D(A.x - B.x, A.y - B.y);
    }
    __device__ Point2D inter(const Point2D &A, const Point2D &B, const Point2D &C, const Point2D &D) { //intersects AB by *the mediator* of CD 
      Point2D M = 0.5*(C+D);
      return A - (dot(A-M, D-C)/dot(B-A, D-C)) * (B-A);
    }

    class Polygon {
    public:
      __device__ Polygon():nbpts(0){};
      __device__ void addPts(Point2D pt) {
        pts[nbpts] = pt;
        nbpts++;
      }; 
      __device__ Polygon& operator=(const Polygon& rhs) {
        nbpts = rhs.nbpts;
        dummy = rhs.dummy;
        memcpy(pts, rhs.pts, nbpts*sizeof(Point2D));
        return *this;
      }
      __device__ void cut(const Point2D &inside_pt, const Point2D &outside_pt) {

        int new_nbpts = 0;
        Point2D newpts[128];
        Point2D AB(outside_pt-inside_pt);
        Point2D M(0.5*(outside_pt+inside_pt));
        double ABM = dot(AB, M);

        Point2D S = pts[nbpts-1];

        for (int i=0; i<nbpts; i++) {

          Point2D E = pts[i];

          double ddot = -ABM + dot(AB, E);
          if (ddot<0) { // E inside clip edge
            double ddot2 = -ABM + dot(AB, S);
            if (ddot2>0) {
               newpts[new_nbpts] = inter(S,E, inside_pt, outside_pt);
               new_nbpts++;
            }
            newpts[new_nbpts] = E;
            new_nbpts++;
          } else {
            double ddot2 = -ABM + dot(AB, S);
            if (ddot2<0) {
               newpts[new_nbpts] = inter(S,E, inside_pt, outside_pt);
               new_nbpts++;
            }       
          }
          S = E;
        }

        memcpy(pts, newpts, min(128, new_nbpts)*sizeof(Point2D));
        nbpts = new_nbpts;
      }

    //private:
     Point2D pts[128];
     int nbpts;
     float dummy;
    };


    __global__ void cut_poly(float *a, Polygon* polygons, int N)
    {
      int idx = blockIdx.x * blockDim.x + threadIdx.x;
      if (idx>=N/2) return;

      Polygon pol;
      pol.addPts(Point2D(0.,0.));
      pol.addPts(Point2D(1.,0.));
      pol.addPts(Point2D(1.,1.));
      pol.addPts(Point2D(0.,1.));

      Point2D curPt(a[2*idx], a[2*idx+1]);

      for (int i=0; i<N/2; i++) {
        Point2D other_pt(a[2*i], a[2*i+1]);
        pol.cut(curPt, other_pt);
      }
      pol.dummy = a[idx];

      polygons[idx] = pol;
    }



    int main(int argc, unsigned char* argv[])
    {

      const int N = 100; 
      float a_h[N], *a_d; 
      Polygon p_h[N/2], *p_d;

      size_t size = N * sizeof(float);
      size_t size_pol = N/2 * sizeof(Polygon);

      cudaError_t err  = cudaMalloc((void **) &a_d, size);   
      cudaError_t err2 = cudaMalloc((void **) &p_d, size_pol);  

      for (int i=0; i<N; i++) a_h[i] = (float)(rand()%1000)*0.001;
      cudaMemcpy(a_d, a_h, size, cudaMemcpyHostToDevice);

      int block_size = 4;
      int n_blocks = N/block_size + (N%block_size == 0 ? 0:1);
      cut_poly <<< n_blocks, block_size >>> (a_d, p_d, N);

      cudaMemcpy(a_h, a_d, sizeof(float)*N, cudaMemcpyDeviceToHost);
      cudaMemcpy(p_h, p_d, sizeof(Polygon)*N/2, cudaMemcpyDeviceToHost);

      for (int i=0; i<N/2; i++)
       printf("%f \t %f \t %u\n", a_h[i], p_h[i].dummy, p_h[i].nbpts);

      cudaFree(a_d);
      cudaFree(p_d);


        return 0;
    }

C++ 中的相同代码可以正常工作：

#include <stdio.h>
#include <iostream>
#include <stdlib.h>

class Point2D {
 public:
 Point2D(double xx=0, double yy=0):x(xx),y(yy){};
 double x, y;
};

double dot(const Point2D &A, const Point2D &B) {
 return A.x*B.x + A.y*B.y;
}
Point2D operator*(double a, const Point2D &P) {
 return Point2D(a*P.x, a*P.y);
}
Point2D operator+(Point2D A, const Point2D &B) {
 return Point2D(A.x + B.x, A.y + B.y);
}
Point2D operator-(Point2D A, const Point2D &B) {
 return Point2D(A.x - B.x, A.y - B.y);
}
Point2D inter(const Point2D &A, const Point2D &B, const Point2D &C, const Point2D &D) { //intersects AB by *the mediator* of CD 
  Point2D M = 0.5*(C+D);
  return A - (dot(A-M, D-C)/dot(B-A, D-C)) * (B-A);
}

class Polygon {
public:
  Polygon():nbpts(0){};
  void addPts(Point2D pt) {
    pts[nbpts] = pt;
    nbpts++;
  }; 
  Polygon& operator=(const Polygon& rhs) {
    nbpts = rhs.nbpts;
    dummy = rhs.dummy;
    memcpy(pts, rhs.pts, nbpts*sizeof(Point2D));
    return *this;
  }
  void cut(const Point2D &inside_pt, const Point2D &outside_pt) {

    int new_nbpts = 0;
    Point2D newpts[128];
    Point2D AB(outside_pt-inside_pt);
    Point2D M(0.5*(outside_pt+inside_pt));
    double ABM = dot(AB, M);

    Point2D S = pts[nbpts-1];

    for (int i=0; i<nbpts; i++) {

      Point2D E = pts[i];

      double ddot = -ABM + dot(AB, E);
      if (ddot<0) { // E inside clip edge
        double ddot2 = -ABM + dot(AB, S);
        if (ddot2>0) {
           newpts[new_nbpts] = inter(S,E, inside_pt, outside_pt);
           new_nbpts++;
        }
        newpts[new_nbpts] = E;
        new_nbpts++;
      } else {
        double ddot2 = -ABM + dot(AB, S);
        if (ddot2<0) {
           newpts[new_nbpts] = inter(S,E, inside_pt, outside_pt);
           new_nbpts++;
        }
      }
        S = E;
    }

    memcpy(pts, newpts, std::min(128, new_nbpts)*sizeof(Point2D));
    /*for (int i=0; i<128; i++) {
      pts[i] = newpts[i];
    }*/
    nbpts = new_nbpts;
  }

//private:
 Point2D pts[128];
 int nbpts;
 float dummy;
};


void cut_poly(int idx, float *a, Polygon* polygons, int N)
{
  if (idx>=N/2) return;

  Polygon pol;
  pol.addPts(Point2D(0.,0.));
  pol.addPts(Point2D(1.,0.));
  pol.addPts(Point2D(1.,1.));
  pol.addPts(Point2D(0.,1.));

  Point2D curPt(a[2*idx], a[2*idx+1]);

  for (int i=0; i<N/2; i++) {
    if (idx==i) continue;
    Point2D other_pt(a[2*i], a[2*i+1]);
    pol.cut(curPt, other_pt);
  }
  pol.dummy = a[idx];

  polygons[idx] = pol;
}



int main(int argc, unsigned char* argv[])
{

  const int N = 100;  // Number of elements in arrays
  float a_h[N], *a_d;  // Pointer to host & device arrays
  Polygon p_h[N/2], *p_d;

  for (int i=0; i<N; i++) a_h[i] = (float)(rand()%1000)*0.001;

  for (int idx=0; idx<N; idx++)
    cut_poly(idx, a_h, p_h, N);

  for (int i=0; i<N/2; i++)
   printf("%f \t %f \t %u\n", a_h[i], p_h[i].dummy, p_h[i].nbpts);

   return 0;
}

Answer 1

好吧，我想你可以无视我的大部分评论。我错误地在使用 CUDA 3.2 设置的机器上工作，并且它在内核启动失败方面表现不同。当我切换到 CUDA 4.1 和 CUDA 5.0 时，事情开始变得有意义。为我在那里的困惑道歉。

无论如何，在过去之后，我很快注意到您的 CPU 和 GPU 实现之间存在差异。具体在这里（查看CPU代码）：

void cut_poly(int idx, float *a, Polygon* polygons, int N)
{
  if (idx>=N/2) return;

  Polygon pol;
  pol.addPts(Point2D(0.,0.));
  pol.addPts(Point2D(1.,0.));
  pol.addPts(Point2D(1.,1.));
  pol.addPts(Point2D(0.,1.));

  Point2D curPt(a[2*idx], a[2*idx+1]);

  for (int i=0; i<N/2; i++) {
    if (idx==i) continue;     /*   NOTE THIS LINE MISSING FROM YOUR GPU CODE */
    Point2D other_pt(a[2*i], a[2*i+1]);
    pol.cut(curPt, other_pt);
  }
  pol.dummy = a[idx];

  polygons[idx] = pol;
}

参考我在上面添加注释的行，如果您将确切的代码行添加到cut_poly内核中 GPU 代码的相应位置，那么无论如何，GPU 代码都会产生与 CPU 代码相同的打印结果。

我要进行的另一项观察是，您不必要地运行只有 4 个线程的块。当您解决代码中的问题时，这没有什么问题，但是一旦您将其用于“生产”目的，您很可能希望将目标设置为更高的数字，例如 256，并确保选择的数字是32 的整数倍，以获得最佳性能。

针对评论中发布的问题，我认为数据已正确复制，但很可能您没有在主机上正确访问它。（我不知道您如何确定“我的数组未正确返回给主机”）。您的大多数类定义都是__device__唯一的。因此，很难访问主机上类中的结构（例如，Point2D pts类中的Polygon类）。我在这里插入修改后的代码，我认为这表明数据正在传输回主机：

    #include <stdio.h>
    #include <iostream>
    #include <stdlib.h>
//    #include <cuda.h>

#define cudaCheckErrors(msg) \
    do { \
        cudaError_t __err = cudaGetLastError(); \
        if (__err != cudaSuccess) { \
            fprintf(stderr, "Fatal error: %s (%s at %s:%d)\n", \
                msg, cudaGetErrorString(__err), \
                __FILE__, __LINE__); \
            fprintf(stderr, "*** FAILED - ABORTING\n"); \
            exit(1); \
        } \
    } while (0)


    class Point2D {
     public:
     __host__ __device__ Point2D(double xx=0, double yy=0):x(xx),y(yy){};
     double x, y;
    };

    __host__ __device__ double dot(const Point2D &A, const Point2D &B) {
     return A.x*B.x + A.y*B.y;
    }
    __host__ __device__ Point2D operator*(double a, const Point2D &P) {
     return Point2D(a*P.x, a*P.y);
    }
    __host__ __device__ Point2D operator+(Point2D A, const Point2D &B) {
     return Point2D(A.x + B.x, A.y + B.y);
    }
    __host__ __device__ Point2D operator-(Point2D A, const Point2D &B) {
     return Point2D(A.x - B.x, A.y - B.y);
    }
    __host__ __device__ Point2D inter(const Point2D &A, const Point2D &B, const Point2D &C, const Point2D &D) { //intersects AB by *the mediator* of CD
      Point2D M = 0.5*(C+D);
      return A - (dot(A-M, D-C)/dot(B-A, D-C)) * (B-A);
    }

    class Polygon {
    public:
      __host__ __device__ Polygon():nbpts(0){};
      __host__ __device__ void addPts(Point2D pt) {
        pts[nbpts] = pt;
        nbpts++;
      };
      __host__ __device__ Polygon& operator=(const Polygon& rhs) {
        nbpts = rhs.nbpts;
        dummy = rhs.dummy;
        memcpy(pts, rhs.pts, nbpts*sizeof(Point2D));
        return *this;
      }
      __host__ __device__ Point2D getpoint(unsigned i){
        if (i<128) return pts[i];
        else return pts[0];
        }
      __host__ __device__ void cut(const Point2D &inside_pt, const Point2D &outside_pt) {

        int new_nbpts = 0;
        Point2D newpts[128];
        Point2D AB(outside_pt-inside_pt);
        Point2D M(0.5*(outside_pt+inside_pt));
        double ABM = dot(AB, M);

        Point2D S = pts[nbpts-1];

        for (int i=0; i<nbpts; i++) {

          Point2D E = pts[i];

          double ddot = -ABM + dot(AB, E);
          if (ddot<0) { // E inside clip edge
            double ddot2 = -ABM + dot(AB, S);
            if (ddot2>0) {
               newpts[new_nbpts] = inter(S,E, inside_pt, outside_pt);
               new_nbpts++;
            }
            newpts[new_nbpts] = E;
            new_nbpts++;
          } else {
            double ddot2 = -ABM + dot(AB, S);
            if (ddot2<0) {
               newpts[new_nbpts] = inter(S,E, inside_pt, outside_pt);
               new_nbpts++;
            }
          }
          S = E;
        }

        memcpy(pts, newpts, min(128, new_nbpts)*sizeof(Point2D));
        nbpts = new_nbpts;
      }

    //private:
     Point2D pts[128];
     int nbpts;
     float dummy;
    };


    __global__ void cut_poly(float *a, Polygon* polygons, int N)
    {
      int idx = blockIdx.x * blockDim.x + threadIdx.x;
      if (idx>=N/2) return;

      Polygon pol;
      pol.addPts(Point2D(0.,0.));
      pol.addPts(Point2D(1.,0.));
      pol.addPts(Point2D(1.,1.));
      pol.addPts(Point2D(0.,1.));

      Point2D curPt(a[2*idx], a[2*idx+1]);

      for (int i=0; i<N/2; i++) {
        if (idx==i) continue;
        Point2D other_pt(a[2*i], a[2*i+1]);
        pol.cut(curPt, other_pt);
      }
      pol.dummy = pol.getpoint(0).x;

      polygons[idx] = pol;
    }



    int main(int argc, unsigned char* argv[])
    {

      const int N = 100;
      float a_h[N], *a_d;
      Polygon p_h[N/2], *p_d;

      size_t size = N * sizeof(float);
      size_t size_pol = N/2 * sizeof(Polygon);

      cudaMalloc((void **) &a_d, size);
      cudaCheckErrors("cm1");
      cudaMalloc((void **) &p_d, size_pol);
      cudaCheckErrors("cm2");

      for (int i=0; i<N; i++) a_h[i] = (float)(rand()%1000)*0.001;
      cudaMemcpy(a_d, a_h, size, cudaMemcpyHostToDevice);
      cudaCheckErrors("cmcp1");

      int block_size = 128;
      int n_blocks = N/block_size + (N%block_size == 0 ? 0:1);
      cut_poly <<< n_blocks, block_size >>> (a_d, p_d, N);
      cudaCheckErrors("kernel");

      cudaMemcpy(a_h, a_d, sizeof(float)*N, cudaMemcpyDeviceToHost);
      cudaCheckErrors("cmcp2");
      cudaMemcpy(p_h, p_d, sizeof(Polygon)*N/2, cudaMemcpyDeviceToHost);
      cudaCheckErrors("cmcp3");

      for (int i=0; i<N/2; i++)
       printf("%f \t %f \t %f \t %u\n", a_h[i], p_h[i].dummy, p_h[i].getpoint(0).x, p_h[i].nbpts);

      cudaFree(a_d);
      cudaFree(p_d);


        return 0;
    }

我建议对这些事情发布新问题。