5. Athread编程常见错误指南

Athread编程常见错误指南


  • 错误一

    代码：

    void fd_compute(float  *a) {

   volatile int get_reply, put_reply;
   float a_slave[10];
   short b_slave[10];

   get_reply = 0;                                                                                                                                 
   athread_get(PE_MODE, a, a_slave,  6,  &get_reply, 0, 0, 0); 
   while (get_reply != 1); 

   get_reply = 0;
   athread_get(PE_MODE, a, a_slave,  0,  &get_reply, 0, 0, 0); 
   while (get_reply != 1); 

   get_reply = 0;
   athread_get(PE_MODE, a, b_slave + 1,  4,  &get_reply, 0, 0, 0); 
   while (get_reply != 1); 


}

    错误现场：

    CG  vn000157 exception !
DMA descriptor examination warning
ERROR, NO SPE EXCEPTION!

    本例错误原因:

    • 传输量没有4B对界(不是4的倍数)
    • 传输量小于0
    • LDM地址没有4B对界，主存地址没对界也会出现该错误

    错误二

    代码：

    void fd_compute(float  *a) {

    volatile int get_reply, put_reply;
    float a_slave[16*1024]; // 16 * 1024 * sizeof(float) = 64KB

   get_reply = 0;
   athread_get(PE_MODE, a, a_slave,  sizeof(float), &get_reply, 0, 0, 0); 
   while (get_reply != 1); 

}

    错误现场：

    The exact shot pc is : 4ff0410178
Slave ldm stack underflow, the stack pointer is now -192. 5B K+�V[+�V["[?1�xTxTThe exact shot pc is : 4ff0410178
Slave ldm stack underflow, the stack pointer is now -192. ��Preserved stack is 0B, You should either:
   1. specify a preserved size more than it.
   Or
   2. get rid of -b 1 option, that will lay slave stack in memory.
TxTPreserved stack is 0B, You should either:

    本例错误原因

    • 定义的数组超出从核 64KB LDM 容量

    错误三

    代码：

    void fd_compute(short  *a) {

    int get_reply, put_reply;                                                                                                                     
    float a_slave[10];

    get_reply = 0;
    athread_get(PE_MODE, a, a_slave,  sizeof(float), &get_reply, 0, 0, 0); 
    while (get_reply != 1); 

}

    错误现场：

    Job <43054411> has been submitted to queue <q_sw_share>
waiting for dispatch ...
dispatching ...

    本例错误原因:

    • 回答字get_reply没有用volatile修饰，可能被编译器优化掉了，然后就卡在while (get_reply != 1)

    错误四

    代码：

    void fd_compute(short  *a) {

    volatile int get_reply, put_reply;
    float a_slave[10];

    get_reply = 0;
    athread_get(PE_MODE, a, a_slave,  10 * sizeof(float), &get_reply, 0, 0, 0); 
    while (get_reply != 1); 

    printf("%f\n",a_slave[100]);
                                                                                                                                                  
}

    错误现场：

    The exact shot pc is : 4ff0410194
SPE 0 exception !
LDM access exception
The exact shot pc is : 4ff0410194
The exact shot pc is : 4ff0410194
SPE 1 exception !
LDM access exception
The exact shot pc is : 4ff0410194
The exact shot pc is : 4ff0410194
SPE 2 exception !

    本例错误原因：

    • 访存越界
    1
    L
     1 条回复

  • 大家可以先用OpenACC熟悉一下编程环境，但如果想取得非常理想的优化效果，建议大家用Athread

    Athread编程出现的错误，一般来源于athread_get/athread_put两个接口，因此在调试过程中可以把程序中其他部分都注释掉，然后耐心研究这两个接口中的参数

    0
    K
     1 条回复
  • 小

    0_1538795623115_8245db8a-7908-414d-8a62-97940102de37-image.png
    请问这个错误的原因是什么呢？怎么解决呢？

    0
    P
     1 条回复
  • K

    @桑薇 请问一下错误一中，主存和LDM地址怎么对界？

    0
    P
     1 条回复
  • P

    @小学生 从核代码里主存地址越界了，检查从核代码里用的主核地址。

    0
  • P

    @kun 一般情况下，浮点数组或整型数组都是四字节对齐的，但是字符数组是一字节对齐的，这种的话，你可能要注意athread的地址参数是不是四字节对齐的。另外数组地址对齐，编译器是可以保证栈申请的数组首地址是32字节对齐的。

    0
    K
     2 条回复
  • K

    此回复已被删除！
    0
  • K

    @popo 好的，谢谢！

    0
  • B

    您好，请问出现这个错误可能是因为什么。。。。。。

    slave.c:16: warning: passing argument 5 of 'athread_get' discards qualifiers from pointer target type
    slave.c:17: warning: passing argument 5 of 'athread_get' discards qualifiers from pointer target type
    slave.c:18: warning: passing argument 5 of 'athread_get' discards qualifiers from pointer target type
    slave.c:19: warning: passing argument 5 of 'athread_get' discards qualifiers from pointer target type
    slave.c:33: warning: passing argument 5 of 'athread_put' discards qualifiers from pointer target type
    slave.c:11: warning: unused variable 'y'
    slave.c:11: warning: unused variable 'x'
    /tmp/ccspin#.qvgvx5.s: Assembler messages:
    /tmp/ccspin#.qvgvx5.s:21: Warning: make sure the slave code is in .text1 section, not .text section
    /tmp/ccspin#.qvgvx5.s:26: Warning: make sure the slave code is in .text1 section, not .text section
    /tmp/ccspin#.qvgvx5.s:77: Error: unknown opcode setfpec0' /tmp/ccspin#.qvgvx5.s:78: Error: unknown opcodesetfpec0'
    /tmp/ccspin#.qvgvx5.s:82: Error: unknown opcode setfpec3' /tmp/ccspin#.qvgvx5.s:90: Error: unknown opcodesetfpec0'
    /tmp/ccspin#.qvgvx5.s:144: Error: unknown opcode setfpec3' /tmp/ccspin#.qvgvx5.s:151: Error: unknown opcodesetfpec0'
    /tmp/ccspin#.qvgvx5.s:181: Error: unknown opcode setfpec3' /tmp/ccspin#.qvgvx5.s:187: Error: unknown opcodesetfpec0'
    /tmp/ccspin#.qvgvx5.s:218: Error: unknown opcode setfpec3' /tmp/ccspin#.qvgvx5.s:224: Error: unknown opcodesetfpec0'
    /tmp/ccspin#.qvgvx5.s:234: Error: unknown opcode setfpec3' /tmp/ccspin#.qvgvx5.s:240: Error: unknown opcodesetfpec0'
    /tmp/ccspin#.qvgvx5.s:367: Error: inappropriate arguments for opcode ldl' /tmp/ccspin#.qvgvx5.s:372: Error: inappropriate arguments for opcodeldl'
    /tmp/ccspin#.qvgvx5.s:376: Error: inappropriate arguments for opcode ldl' /tmp/ccspin#.qvgvx5.s:380: Error: inappropriate arguments for opcodeldl'
    /tmp/ccspin#.qvgvx5.s:385: Error: inappropriate arguments for opcode ldl' /tmp/ccspin#.qvgvx5.s:389: Error: inappropriate arguments for opcodeldl'
    /tmp/ccspin#.qvgvx5.s:393: Error: inappropriate arguments for opcode ldl' /tmp/ccspin#.qvgvx5.s:403: Error: inappropriate arguments for opcodestl'
    /tmp/ccspin#.qvgvx5.s:459: Error: unknown opcode setfpec3' /tmp/ccspin#.qvgvx5.s:466: Error: unknown opcodesetfpec0'
    /tmp/ccspin#.qvgvx5.s:496: Error: unknown opcode `setfpec3'
    /tmp/ccspin#.qvgvx5.s:536: Warning: make sure the slave code is in .text1 section, not .text section

    0 1 条回复
  • D

    此回复已被删除！
    0
登录后回复