CUDA on Thrust：如何实现优先级队列
我的计划是使用皮尔逊相关性计算距离矩阵，并从距离矩阵中获取每个节点的 q- 最近邻 (q=ln(n)) 并将它们放入结果向量中。我在 C++ 中使用相关函数循环内的 STL 优先级队列来完成此操作。
但你认为，有什么办法可以在 GPU 中做到这一点吗？
- 任何人都可以帮助我，我怎样才能在GPU中做同样的事情（可能推力对我来说会更容易！）
- 如何在GPU中实现优先级队列？
这是我的 CPU(C++ STL) 代码：
例如，
distance matrix
-----------------------
0 3 2 4
3 0 4 5
2 4 0 6
.....
output in a object vector
=================
source target weight
--------------------------------
0 2 2
0 1 3 .... (sorted by Edge weight)
1 0 3
1 2 4
2 0 2
.....
calculatePearsonCorrelation(float vector1[], float vector2[], int m) {
// float distancePearson(vector vector1, vector vector2){
int i;
float a=0,b=0,c=0,d=0, e = 0, sumX=0, sumY=0;
//m = vector1.size();
for(i=0;iq){
MIN=pqx.top().get_corr();
if(corr::iterator it = qNNVector.begin(); it!=qNNVector.end(); ++it) {
fout
my plan is to calculate a distance matrix using Pearsons's correlation and get q- nearest neighbors for each nodes (q=ln(n)) from the distance matrix and put them in a result vector. I did it in C++ using STL priority queue inside the loop of correlation function.
But do you think, there is any way to do it in GPU?
- Can anyone help me, how can I do the same in GPU (probably Thrust will be easier for me!)
- How to implement priority queue in GPU?
Here is my CPU(C++ STL) code:
For example,
distance matrix
-----------------------
0 3 2 4
3 0 4 5
2 4 0 6
.....
output in a object vector
=================
source target weight
--------------------------------
0 2 2
0 1 3 .... (sorted by Edge weight)
1 0 3
1 2 4
2 0 2
.....
calculatePearsonCorrelation(float vector1[], float vector2[], int m) {
// float distancePearson(vector vector1, vector vector2){
int i;
float a=0,b=0,c=0,d=0, e = 0, sumX=0, sumY=0;
//m = vector1.size();
for(i=0;iq){
MIN=pqx.top().get_corr();
if(corr::iterator it = qNNVector.begin(); it!=qNNVector.end(); ++it) {
fout
如果你对这篇内容有疑问，欢迎到本站社区发帖提问 参与讨论，获取更多帮助，或者扫码二维码加入 Web 技术交流群。
绑定邮箱获取回复消息
由于您还没有绑定你的真实邮箱，如果其他用户或者作者回复了您的评论，将不能在第一时间通知您！
发布评论
评论（1）
您是否尝试过使用 Thrust 排序？您只需读出前 q 个元素即可。如果有足够的元素，您可能会看到相当大的加速。
Have you tried using Thrust's sort? You can just read out the first q elements. If there are enough elements, you'll likely see a decent speedup.