如何使用推力中的映射选择的点的两个device_vector的加权平均值?

时间:2018-11-23 13:19:42

标签: cuda thrust weighted-average

我有两个device_vector P&Q(假设大小为100)。 我有两个用于P&Q的device_vector映射(MapP和MapQ的大小为10),其中包含要从P&Q中选择的点的索引。 我有一个device_vector D作为重量。

我需要为P&Q中使用相应地图选择的所有点计算(P * D + Q)/(D + 1)。

我的方法如下。它可以工作,但是太麻烦了。 有人可以建议一种更好的方法吗?

#include <thrust/device_vector.h>
#include <thrust/random.h>
#include <thrust/sequence.h>
#include <thrust/execution_policy.h>
#include <thrust/transform.h>
#include <thrust/functional.h>
#include <thrust/iterator/transform_iterator.h>
#include <thrust/iterator/permutation_iterator.h>

thrust::device_vector<float> random_vector(const size_t N, 
                                         unsigned int seed = thrust::default_random_engine::default_seed)
{
    thrust::default_random_engine rng(seed);
    thrust::uniform_real_distribution<float> u01(0.0f, 10.0f);
    thrust::device_vector<float> temp(N);
    for(size_t i = 0; i < N; i++) {
        temp[i] = u01(rng);
    }
    return temp;
}

// note: functor inherits from unary_function
struct increment : public thrust::unary_function<int,int>
{
  __host__ __device__
  int operator()(int x) const
  {
    return x + 1;
  }
};

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

int N=atoi(argv[1]);

thrust::device_vector<float> P = random_vector(N,1);
thrust::device_vector<float> Q = random_vector(N,9);

thrust::device_vector<int> D(N);
thrust::sequence(thrust::device, D.begin(), D.begin() + N, 1);

thrust::device_vector<float> temp(10);

thrust::device_vector<int> MapP(10); // map
thrust::device_vector<int> MapQ(10); // map

MapP[0]=0;MapP[1]=5;MapP[2]=4;MapP[3]=2;MapP[4]=7;MapP[5]=1;MapP[6]=9;MapP[7]=3;MapP[8]=6;MapP[9]=8;
MapQ[0]=10;MapQ[1]=15;MapQ[2]=12;MapQ[3]=14;MapQ[4]=11;MapQ[5]=17;MapQ[6]=13;MapQ[7]=19;MapQ[8]=18;MapQ[9]=16;


// The weighted average is (D*P+Q)/(D+1)
// We compute D*P first

//thrust::transform(thrust::device, P.begin(), P.end(), D.begin(), temp.begin(), thrust::multiplies<float>()); // use permutation iterator

thrust::transform(thrust::device, thrust::make_permutation_iterator(P.begin(),MapP.begin()),
                                  thrust::make_permutation_iterator(P.end(),MapP.end()),
                  thrust::make_permutation_iterator(D.begin(),MapP.begin()), 
                  temp.begin(), thrust::multiplies<float>());


// Then we add D*p to Q

//thrust::transform(thrust::device, temp.begin(), temp.end(), Q.begin(), temp.begin(), thrust::plus<float>()); // use permutation iterator

thrust::transform(thrust::device, temp.begin(), temp.end(),
                  thrust::make_permutation_iterator(Q.begin(),MapQ.begin()), 
                  temp.begin(), thrust::plus<float>());


// Then we divide by D+1

//thrust::transform(thrust::device, temp.begin(), temp.end(), thrust::make_transform_iterator(D.begin(), increment()), temp.begin(),  thrust::divides<float>());

thrust::transform(thrust::device, temp.begin(), temp.end(),
                  thrust::make_permutation_iterator(D.begin(),MapP.begin()), 
                  temp.begin(), thrust::divides<float>());


// replace contents of P with the weighted sum using pts in map M

thrust::copy(thrust::device, temp.begin(), temp.end(), thrust::make_permutation_iterator(P.begin(),MapP.begin())); // use permutation iterator

return 0;
}

1 个答案:

答案 0 :(得分:2)

我假设您要对向量进行逐元素运算,因为那是您提供的演示程序代码的行为。

请注意,当传递置换迭代器的末尾时,我们不使用源向量的末尾:

thrust::make_permutation_iterator(P.end(),MapP.end()),
                                  ^^^^^

而是开始:

thrust::make_permutation_iterator(P.begin(),MapP.end()),

有关示例,请参考thrust quick start guide

还请注意,在您的问题和代码中,您指的都是除以D + 1,但是您的代码实际上是除以D,而不是D + 1。

关于您的问题,只需使用适当定义的函子调用thrust::transform即可完成所有操作。由于在此实现中我们需要将多个向量传递给thrust::transform,因此我们将介绍thrust::zip_iterator

$ cat t332.cu
#include <thrust/device_vector.h>
#include <thrust/random.h>
#include <thrust/sequence.h>
#include <thrust/execution_policy.h>
#include <thrust/transform.h>
#include <thrust/functional.h>
#include <thrust/iterator/transform_iterator.h>
#include <thrust/iterator/permutation_iterator.h>
#include <thrust/iterator/zip_iterator.h>

thrust::device_vector<float> random_vector(const size_t N,
                                         unsigned int seed = thrust::default_random_engine::default_seed)
{
    thrust::default_random_engine rng(seed);
    thrust::uniform_real_distribution<float> u01(0.0f, 10.0f);
    thrust::device_vector<float> temp(N);
    for(size_t i = 0; i < N; i++) {
        temp[i] = u01(rng);
    }
    return temp;
}

// The weighted average is (D*P+Q)/(D+1)
struct w_avg
{
template <typename T>
  __host__ __device__
  float operator()(T x) const
  {
    return (thrust::get<0>(x)*thrust::get<1>(x)+thrust::get<2>(x))/(thrust::get<1>(x)+1.0f);
  }
};

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

int N=atoi(argv[1]);

thrust::device_vector<float> P = random_vector(N,1);
thrust::device_vector<float> Q = random_vector(N,9);

thrust::device_vector<int> D(N);
thrust::sequence(thrust::device, D.begin(), D.begin() + N, 1);


thrust::device_vector<int> MapP(10); // map
thrust::device_vector<int> MapQ(10); // map

MapP[0]=0;MapP[1]=5;MapP[2]=4;MapP[3]=2;MapP[4]=7;MapP[5]=1;MapP[6]=9;MapP[7]=3;MapP[8]=6;MapP[9]=8;
MapQ[0]=10;MapQ[1]=15;MapQ[2]=12;MapQ[3]=14;MapQ[4]=11;MapQ[5]=17;MapQ[6]=13;MapQ[7]=19;MapQ[8]=18;MapQ[9]=16;


// The weighted average is (D*P+Q)/(D+1)

thrust::transform(thrust::device, thrust::make_zip_iterator(thrust::make_tuple(
                                                            thrust::make_permutation_iterator(P.begin(),MapP.begin()),
                                                            thrust::make_permutation_iterator(D.begin(),MapP.begin()),
                                                            thrust::make_permutation_iterator(Q.begin(),MapQ.begin()))),
                                  thrust::make_zip_iterator(thrust::make_tuple(
                                                            thrust::make_permutation_iterator(P.begin(),MapP.end()),
                                                            thrust::make_permutation_iterator(D.begin(),MapP.end()),
                                                            thrust::make_permutation_iterator(Q.begin(),MapQ.end()))),
                                                            thrust::make_permutation_iterator(P.begin(),MapP.begin()),
                                  w_avg());


for (int i = 0; i < 5; i++) {
  std::cout << P[i] << std::endl;}
return 0;
}
$ nvcc -o t332 t332.cu
$ ./t332 100
4.02976
3.75275
5.32832
8.53189
8.46641
$

请注意,以上代码中的函子除以D + 1。将其更改为除以D,以匹配您的代码(而不是您声明的意图),将是微不足道的。