如何在opencl中使用公用变量

Question

我正在尝试在pyopencl中实现广度优先搜索。

因此，处理单个工作项很顺利，但是当我尝试通过为每个工作项提供不同的要点搜索来加快速度时，使它们中的一个完成搜索时，它可以停止所有其他工作项并节省时间。我找不到一种使工作项相互交流的方法。

import pyopencl as cl
import numpy as np
import os

TASKS = 0

os.environ['PYOPENCL_COMPILER_OUTPUT'] = '1'
os.environ['PYOPENCL_CTX'] = '1'

# platform, device, context & queue setup
platform = cl.get_platforms()[0]
device = platform.get_devices()[0]
ctx = cl.Context([device])
queue = cl.CommandQueue(ctx)

# graph attributes
keys = np.array([0, 2, 3, 4, 5, 6, 7]) # vertex ids
keys = keys.astype(np.int32)
graph = np.array([
    2, 7, -1,
    0, 3, -1,
    2, 6, -1,
    5, 6, 7, 
    4, 6, 7, 
    3, 4, -1,
    0, 5, -1
])
graph = graph.astype(np.int32)
neighbours = np.array([2, 2, 2, 3, 3, 2, 2]) # of neighbours of each vertex
neighbours = neighbours.astype(np.int32)

visited = np.empty(keys.size)      # visited vertices
frontier = np.full(keys.size, -1)  # next in line to be traversed with arbitrary source
src = 4
goal = np.array([6])               # vertex to find
vertices = np.array([7])           # number of vertices

# dividing task among neighbours of starting vertex
threads = []
first_neighbours = []
for g in range(0, vertices[0]):
    if keys[g] == src:
        for j in range(0, neighbours[g]):
            threads.append(graph[g*3 + j])
        threads = np.array(threads)
        threads = threads.astype(np.int32)
        break
TASKS = len(threads)

# buffer allocation
mf = cl.mem_flags
graph_buf = cl.Buffer(ctx, mf.READ_ONLY | mf.COPY_HOST_PTR, hostbuf=graph)
keys_buf = cl.Buffer(ctx, mf.READ_ONLY | mf.COPY_HOST_PTR, hostbuf=keys)
neighbours_buf = cl.Buffer(ctx, mf.READ_ONLY | mf.COPY_HOST_PTR, hostbuf=neighbours)
visited_buf = cl.Buffer(ctx, mf.READ_WRITE | mf.COPY_HOST_PTR, hostbuf=visited)
frontier_buf = cl.Buffer(ctx, mf.READ_WRITE | mf.COPY_HOST_PTR, hostbuf=frontier)
goal_buf = cl.Buffer(ctx, mf.READ_ONLY| mf.COPY_HOST_PTR, hostbuf=goal)
vertices_buf = cl.Buffer(ctx, mf.READ_ONLY| mf.COPY_HOST_PTR, hostbuf=vertices)
threads_buf = cl.Buffer(ctx, mf.READ_ONLY| mf.COPY_HOST_PTR, hostbuf=threads)

# kernel program
prg = cl.Program(ctx, '''
    __kernel void bfs (__global int graph[], __global int keys[], __global int neighbours[], __global int visited[], __global int frontier[], __global int goal[], __global int vertices[], __global int threads[]) {
        int gid = get_global_id(0);
        frontier[0] = threads[gid];
        int fsize = 1;
        int vsize = 0;
        while (true) {
            // printf("%d Frontier:", gid);
            for (int i=0; i<fsize; i++) {
            //    printf(" %d", frontier[i]);
            }
            // printf("\\n");
            if (fsize == 0) {
                printf("Not found %d\\n", goal[0]);
                return;
            }
            int current = *(frontier + 0);
            // printf("Current: %d\\n", current);
            if (current == goal[0]) {
                printf("Found %d\\n", current);
                return;
            }
            for (int fr=0; fr<fsize-1; fr++) {
                frontier[fr] = frontier[fr+1];
            }
            *(frontier + fsize - 1) = -1;
            fsize--;
            *(visited + vsize) = current;
            vsize++;
            // printf("Visited:");
            for (int i=0; i<vsize; i++) {
                // printf(" %d", visited[i]);
            }
            // printf("\\n");

            for (int s=0; s<vertices[0]; s++) {
                if (keys[s] == current) {
                    // printf("Neighbours: %d:", neighbours[s]);
                    for (int n=0; n<neighbours[s]; n++)  {
                        int nb = graph[s*3 + n];
                        int already = 0;
                        // printf(" %d", nb);
                        for(int v=0; v<vsize; v++) {
                            if (nb == *(visited+v)) {
                                already = 1;
                                break;
                            }
                        }
                        if (already == 1) {
                            continue;
                        }
                        for (int f=0; f<fsize; f++) {
                            if (nb == *(frontier+f)) {
                                already = 1;
                                break;
                            }
                        }
                        if (already == 0) {
                            *(frontier + fsize) = nb;
                            fsize++;
                        }
                    }
                    // printf("\\n");
                    break;
                }
            }
        }
    }
''').build()

prg.bfs(queue, (TASKS,), (1,), graph_buf, keys_buf, neighbours_buf, visited_buf, frontier_buf, goal_buf, vertices_buf, threads_buf)

我知道在搜索成功时，通过打印的消息，每个工作项都将进行整个搜索。他们中只有一个应该这样做。仅当我能找到一种方法来停止所有其他工作项目时。预先感谢。