@tat0:您也可以使用Arrayfire:使用 lower_bound() 的矢量化搜索不会立即给您答案,而在 arrayfire 中使用 setintersect() ,您可以直接获得两个数组的“交集”:
float A_host[] = {3,22,4,5,2,9,234,11,6,17,7,873,23,45,454};
int szA = sizeof(A_host) / sizeof(float);
float B_host[] = {345,5,55,6,7,8,19,2,63};
int szB = sizeof(B_host) / sizeof(float);
// initialize arrays from host data
array A(szA, 1, A_host);
array B(szB, 1, B_host);
array U = setintersect(A, B); // compute intersection of 2 arrays
int n_common = U.elements();
std::cout << "common: ";
print(U);
输出为:常见:U = 2.0000 5.0000 6.0000 7.0000
要获取数组 A 中这些元素的实际位置,可以使用以下构造(前提是 A 中的元素是唯一的):
int n_common = U.elements();
array loc = zeros(n_common); // empty array
gfor(array i, n_common) // parallel for loop
loc(i) = sum((A == U(i))*seq(szA));
print(loc);
然后: loc = 4.0000 3.0000 8.0000 10.0000
此外,thrust::lower_bound() 似乎比 setintersect() 慢,我使用以下程序对其进行了基准测试:
int *g_data = 0;
int g_N = 0;
void thrust_test() {
thrust::device_ptr<int> A = thrust::device_pointer_cast((int *)g_data),
B = thrust::device_pointer_cast((int *)g_data + g_N);
thrust::device_vector<int> output(g_N);
thrust::lower_bound(A, A + g_N, B, B + g_N,
output.begin(),
thrust::less<int>());
std::cout << "thrust: " << output.size() << "\n";
}
void af_test()
{
array A(g_N, 1, g_data, afDevicePointer);
array B(g_N, 1, g_data + g_N, afDevicePointer);
array U = setintersect(A, B);
std::cout << "intersection sz: " << U.elements() << "\n";
}
int main()
{
g_N = 3e6; // 3M entries
thrust::host_vector< int > input(g_N*2);
for(int i = 0; i < g_N*2; i++) { // generate some input
if(i & 1)
input[i] = (i*i) % 1131;
else
input[i] = (i*i*i-1) % 1223 ;
}
thrust::device_vector< int > dev_input = input;
// sort the vector A
thrust::sort(dev_input.begin(), dev_input.begin() + g_N);
// sort the vector B
thrust::sort(dev_input.begin() + g_N, dev_input.begin() + g_N*2);
g_data = thrust::raw_pointer_cast(dev_input.data());
try {
info();
printf("thrust: %.5f seconds\n", timeit(thrust_test));
printf("af: %.5f seconds\n", timeit(af_test));
} catch (af::exception& e) {
fprintf(stderr, "%s\n", e.what());
}
return 0;
}
结果:
CUDA 工具包 4.2,驱动程序 295.59
GPU0 GeForce GT 650M,2048 MB,Compute 3.0(单、双)
内存使用:1937 MB 可用空间(总共 2048 MB)
推力:0.13008 秒
阵列火:0.06702 秒