1

是否可以在 CUDA 中使用 CUDPP 进行分段排序?通过分段排序,我的意思是对受以下标志保护的数组元素进行排序。

A[10,9,8,7,6,5,4,3,2,1]

Flag array[1,0,1,0,0,1,0,0,0,0]

对 A 中连续 1 之间的元素进行排序。
预期输出

[9,10,6,7,8,1,2,3,4,5]
4

1 回答 1

6

您可以在单个排序过程中执行此操作:想法是调整数组中的元素,以便排序仅在“段”内重新定位元素

对于您的示例:

A[10,9,8,7,6,5,4,3,2,1]
flag[0,0,1,0,0,1,0,0,0,0] 

(我删除了第一个 1,因为它不需要)

首先扫描标志数组:

scanned_flag[0,0,1,1,1,2,2,2,2,2]

那么你有很多选项取决于数字类型,例如对于无符号整数,你可以设置最高位来区分“段”。最简单的方法就是将最大元素乘以scanned_flags:

A + scanned_flag*10 = [10,9,18,17,16,25,24,23,22,21]

剩下的很简单:对数组进行排序并反转转换。这是两个版本:使用 Arrayfire 和推力。检查您更喜欢哪个。

阵列火:

void af_test() {
 int A[] = {10,9,8,7,6,5,4,3,2,1};  
 int S[] = {0, 0,1,0,0,1,0,0,0,0};  
 int n = sizeof(A) / sizeof(int);

 af::array devA(n, A, af::afHost);
 af::array devS(n, S, af::afHost);
 // obtain the max element
 int maxi = af::max< int >(devS);

 // scan the keys
 // keys = 0,0,1,1,1,2,2,2,2,2
 af::array keys = af::accum(devS);

 // compute: A = A + keys * maxi
 // A = 10,9,18,17,16,25,24,23,22,21
 devA = devA + keys * maxi;

 // sort the array
 // A = 9,10,16,17,18,21,22,23,24,25
 devA = af::sort(devA);

 // compute: A = A - keys * maxi
 // A = 9,10,6,7,8,1,2,3,4,5
 devA = devA - keys * maxi;
 // print the results
 print(devA);
}

推力:

template<typename T>
struct add_mul : public binary_function<T,T,T>
{
add_mul(const T& _factor) : factor(_factor) {
}
   __host__ __device__ T operator()(const T& a, const T& b) const 
{
    return (a + b * factor);
}
const T factor;
}; 

void thrust_test()
{
  int A[] = {10,9,8,7,6,5,4,3,2,1};  
  int S[] = {0, 0,1,0,0,1,0,0,0,0};  
  int n = sizeof(A) / sizeof(int);
  thrust::host_vector< int > hA(A, A + n), hS(S, S + n);

  thrust::device_vector< int > devA = hA, devS = hS, keys(n);
  // scan the keys 
  thrust::inclusive_scan(devS.begin(), devS.end(), keys.begin());
  // obtain the maximal element
  int maxi = *(thrust::max_element(devA.begin(), devA.end()));
  // compute: A = A + keys * maxi
  thrust::transform(devA.begin(), devA.end(), keys.begin(), devA.begin(), add_mul< int >(maxi)); 
  // sort the array
  thrust::sort(devA.begin(), devA.end());
  // compute: A = A - keys * maxi
  thrust::transform(devA.begin(), devA.end(), keys.begin(), devA.begin(), add_mul< int >(-maxi)); 
  // copy back to the host 
  hA = devA;
  std::cout << "\nSorted array\n";
  thrust::copy(hA.begin(), hA.end(), std::ostream_iterator<int>(std::cout, "\n"));
}
于 2012-11-22T17:33:26.610 回答