使用带步幅的推力计数迭代器

Question

我正在寻找一种使用 thrust::counting_iterator 函数的方法，以便并行化以下 for 循环：

for (int stride = 0 ; stride < N * M ; stride+=M) // N iterations
{
    // Body of the loop
}

代码如下：

struct functor ()
{
   __host__ __device__ void operator() (const int i)
   {
      // Body of the loop
   }
}

thrust::counting_iterator<int> it1(0);
thrust::counting_iterator<int> it2 = it1 + N * M;
thrust::for_each (it1 , it2 , functor());

我知道counting_iterator 将迭代器递增 1，那么有没有办法以 M 递增？

Answer 1

这是arbitrary transformation example 和strided range example 的组合。

下面，我正在考虑一个转换的例子

D[i] = A[i] + B[i] * C[i]

代码如下：

#include <thrust/for_each.h>
#include <thrust/host_vector.h>
#include <thrust/device_vector.h>
#include <thrust/iterator/zip_iterator.h>
#include <iostream>

#include <thrust/iterator/counting_iterator.h>
#include <thrust/iterator/transform_iterator.h>
#include <thrust/iterator/permutation_iterator.h>
#include <thrust/functional.h>

#include <thrust/fill.h>

// for printing
#include <thrust/copy.h>
#include <ostream>

#define STRIDE 2

template <typename Iterator>
class strided_range
{
    public:

    typedef typename thrust::iterator_difference<Iterator>::type difference_type;

    struct stride_functor : public thrust::unary_function<difference_type,difference_type>
    {
        difference_type stride;

        stride_functor(difference_type stride)
            : stride(stride) {}

        __host__ __device__
        difference_type operator()(const difference_type& i) const
        {
            return stride * i;
        }
    };

    typedef typename thrust::counting_iterator<difference_type>                   CountingIterator;
    typedef typename thrust::transform_iterator<stride_functor, CountingIterator> TransformIterator;
    typedef typename thrust::permutation_iterator<Iterator,TransformIterator>     PermutationIterator;

    // type of the strided_range iterator
    typedef PermutationIterator iterator;

    // construct strided_range for the range [first,last)
    strided_range(Iterator first, Iterator last, difference_type stride)
        : first(first), last(last), stride(stride) {}

    iterator begin(void) const
    {
        return PermutationIterator(first, TransformIterator(CountingIterator(0), stride_functor(stride)));
    }

    iterator end(void) const
    {
        return begin() + ((last - first) + (stride - 1)) / stride;
    }

    protected:
    Iterator first;
    Iterator last;
    difference_type stride;
};

struct arbitrary_functor
{
    template <typename Tuple>
    __host__ __device__
    void operator()(Tuple t)
    {
        // D[i] = A[i] + B[i] * C[i];
        thrust::get<3>(t) = thrust::get<0>(t) + thrust::get<1>(t) * thrust::get<2>(t);
    }
};


int main(void)
{
    // allocate storage
    thrust::device_vector<float> A(5);
    thrust::device_vector<float> B(5);
    thrust::device_vector<float> C(5);
    thrust::device_vector<float> D(5);

    // initialize input vectors
    A[0] = 3;  B[0] = 6;  C[0] = 2; 
    A[1] = 4;  B[1] = 7;  C[1] = 5; 
    A[2] = 0;  B[2] = 2;  C[2] = 7; 
    A[3] = 8;  B[3] = 1;  C[3] = 4; 
    A[4] = 2;  B[4] = 8;  C[4] = 3; 

    typedef thrust::device_vector<float>::iterator Iterator;
    strided_range<Iterator> posA(A.begin(), A.end(), STRIDE);
    strided_range<Iterator> posB(B.begin(), B.end(), STRIDE);
    strided_range<Iterator> posC(C.begin(), C.end(), STRIDE);
    strided_range<Iterator> posD(D.begin(), D.end(), STRIDE);

    // apply the transformation
    thrust::for_each(thrust::make_zip_iterator(thrust::make_tuple(posA.begin(), posB.begin(), posC.begin(), posD.begin())),
                 thrust::make_zip_iterator(thrust::make_tuple(posA.end(), posB.end(), posC.end(), posD.end())),
                 arbitrary_functor());

    // print the output
    for(int i = 0; i < 5; i++)
    std::cout << A[i] << " + " << B[i] << " * " << C[i] << " = " << D[i] << std::endl;
}

Answer 2

为什么不在仿函数中将i 变量乘以M？

如果M 在编译时已知，它可能是：

struct functor 
{
   __host__ __device__ void operator() (const int my_i)
   {
      int i = my_i *M;
      // Body of the loop
   }
};

thrust::counting_iterator<int> it1(0);
thrust::counting_iterator<int> it2 = it1 + N;
thrust::for_each (it1 , it2 , functor());

如果M 仅在运行时已知，我们可以将其作为初始化参数传递给函子：

struct functor 
{
   int my_M;
   functor(int _M) : my_M(_M) ();
   __host__ __device__ void operator() (const int my_i)
   {
      int i = my_i *my_M;
      // Body of the loop
   }
};

thrust::counting_iterator<int> it1(0);
thrust::counting_iterator<int> it2 = it1 + N;
thrust::for_each (it1 , it2 , functor(M));

您还可以将计数迭代器包装在转换迭代器中，它接受计数迭代器并将其乘以 M：

struct functor 
{
   __host__ __device__ void operator() (const int i)
   {
      // Body of the loop
   }
};

using namespace thrust::placeholders;
thrust::counting_iterator<int> it1(0);
thrust::counting_iterator<int> it2 = it1 + N;
thrust::for_each (make_transform_iterator(it1, _1 * M) , thrust::make_transform_iterator(it2, _1 * M) , functor());

最后一个示例使用thrust placeholder expressions，尽管它可以通过一个额外的普通函子等效地实现，该函子返回其参数乘以它的参数。

这是一个完整的示例，展示了所有 3 种方法：

$ cat t492.cu
#include <stdio.h>
#include <thrust/transform.h>
#include <thrust/for_each.h>
#include <thrust/execution_policy.h>
#include <thrust/iterator/counting_iterator.h>
#include <thrust/iterator/transform_iterator.h>
#include <thrust/host_vector.h>
#include <thrust/functional.h>
#define N 5
#define M 4
using namespace thrust::placeholders;

struct my_functor_1
{
  __host__ __device__  void operator() (const int i)
  {
    printf("functor 1 value: %d\n", i);
  }
};

struct my_functor_2
{
   __host__ __device__ void operator() (const int my_i)
   {
    int i = my_i*M;
    printf("functor 2 value: %d\n", i);
   }
};

struct my_functor_3
{
   int my_M;
   my_functor_3(int _M) : my_M(_M) {};
   __host__ __device__ void operator() (const int my_i)
   {
      int i = my_i *my_M;
      printf("functor 3 value: %d\n", i);
   }
};


int main(){
  thrust::counting_iterator<int> it1(0);
  thrust::counting_iterator<int> it2 = it1 + N;
  thrust::for_each(thrust::host, it1, it2, my_functor_1());
  thrust::for_each(thrust::host, it1, it2, my_functor_2());
  thrust::for_each(thrust::host, it1, it2, my_functor_3(M));
  thrust::for_each(thrust::host, thrust::make_transform_iterator(it1, _1 * M), thrust::make_transform_iterator(it2, _1 * M), my_functor_1());
  return 0;
}


$ nvcc -arch=sm_20 -o t492 t492.cu
$ ./t492
functor 1 value: 0
functor 1 value: 1
functor 1 value: 2
functor 1 value: 3
functor 1 value: 4
functor 2 value: 0
functor 2 value: 4
functor 2 value: 8
functor 2 value: 12
functor 2 value: 16
functor 3 value: 0
functor 3 value: 4
functor 3 value: 8
functor 3 value: 12
functor 3 value: 16
functor 1 value: 0
functor 1 value: 4
functor 1 value: 8
functor 1 value: 12
functor 1 value: 16
$