Algorithm 1 CUDA_BFS(Graph G(V, E), Source Vertex S)

{

1. Create vertex array Va from all vertices in G(V, E);

2. Create edge array Ea from all edges in G(V, E);

3. Create Frontier array Fa, visited array Xa and cost array Ca of size V;

4. Initialize Fa, Xa to false and Ca to INFINITE;

5. Fa[s]=true;

6. Ca[s]=0;

7. While Fa not Empty do

8.      for each vertex V in parallel do

9.            Invoke CUDA_BFS_KERNEL(Va, Ea, Fa, Xa, Ca) on the grid.

10.    end for

11. end while

}

Algorithm 2 CUDA_BFS_KERNEL(Va, Ea, Fa, Xa, Ca)

{

1. tid=getThreadID;

2. if Fa[tid] then

3.    Fa[tid]=false;Xa[tid]=true;

4.    for all neighbors nid of tid do

5.          if NOT Xa[nid] then

6.             Ca[nid]=Ca[nid]+1

7.             Fa[nid]=true;

8.          end if

9.    end for

10.end if

}

这个算法

Algorithm 1 CUDA_BFS(Graph G(V, E), Source Vertex S)

{

1. Create vertex array Va from all vertices in G(V, E);

2. Create edge array Ea from all edges in G(V, E);

3. Create Frontier array Fa, visited array Xa and cost array Ca of size V;

4. Initialize Fa, Xa to false and Ca to INFINITE;

5. Fa[s]=true;

6. Ca[s]=0;

7. While Fa not Empty do

8.      for each vertex V in parallel do

9.            Invoke CUDA_BFS_KERNEL(Va, Ea, Fa, Xa, Ca) on the grid.

10.    end for

11. end while

}

Algorithm 2 CUDA_BFS_KERNEL(Va, Ea, Fa, Xa, Ca)

{

1. tid=getThreadID;

2. if Fa[tid] then

3.    Fa[tid]=false;Xa[tid]=true;

4.    for all neighbors nid of tid do

5.          if NOT Xa[nid] then

6.             Ca[nid]=Ca[nid]+1

7.             Fa[nid]=true;

8.          end if

9.    end for

10.end if

}

这个算法实现了level synchronization， 一个顶点一个线程，算法复杂度是图的直径（就是从源到叶子的最长距离）即O(diameter).

其中数据结构的设计和存储上可以做很多优化。。。这图要是动态地根据一定规则自动生成，那么数据结构的动态性会影响到性能么？多大程度上？问问自己，动动手动动脑子~

学习自：Accelerating large graph algorithms on the GPU using CUDA

文章来源：http://blog.csdn.net/iJuliet/archive/2009/03/09/3969391.aspx