使用 OpenCL 管道时内核挂起

Question

我正在尝试编写一个使用 OpenCL 管道的 OpenCL 内核。内核代码如下。

    uint tid = get_global_id(0);
    uint numWorkItems = get_global_size(0);
    int i;
    int rid;
    int temp = 0, temp1 = 0;
    int val;

    int    szgr = get_local_size(0);
    int    lid  = get_local_id(0);

    for(i = tid + start_index; i < rLen; i = i + numWorkItems){
            temp = 0;
            val = input[i];
            temp = hashTable[val - 1];
            if(temp){
                    temp1 = projection[val - 1];
            }

            reserve_id_t rid1 = work_group_reserve_write_pipe(c0, szgr);

            while(is_valid_reserve_id(rid1) == false){
            rid1 = work_group_reserve_write_pipe(c0, szgr);
            }

            if(is_valid_reserve_id(rid1))
            {

                    write_pipe(c0,rid1,lid, &temp);
                    work_group_commit_write_pipe(c0, rid1);
            }

            reserve_id_t rid2 = work_group_reserve_write_pipe(c1, szgr);

            while(is_valid_reserve_id(rid2) == false){
            rid2 = work_group_reserve_write_pipe(c1, szgr);
            }

            if(is_valid_reserve_id(rid2))
            {

                    write_pipe(c1,rid2,lid, &temp1);
                    work_group_commit_write_pipe(c1, rid2);
            }
    }

但是 work_group_reserve_write_pipe 函数总是失败，因此内核在 while 循环中挂起。如果我删除这个 while 循环，那么代码不会挂起，但不会写入管道。谁能告诉我为什么会这样？

管道被声明为 _write_only 管道。

Answer 1

About work_group_reserve_write_pipe:

This built-in function must be encountered by all work-items in a work-group executing the kernel with the same argument values; otherwise the behavior is undefined.

循环从 tid + start_index 开始，因此在一些循环迭代之后，一些工作项没有命中此指令。还有一个 while 循环正在做同样的未定义行为。

使用 OpenCL 管道时内核挂起

Kernel hangs when using OpenCL pipes

opencl

amd-gpu

opencl-pipes