Posts made by laser2000

复杂访问大概是，数组元素访问编号不是循环变量，而是由循环变量计算后得到。
这只是一个警告，如果你确定访问没有问题，就不管它

查看中间代码

void slave_entire_6(void *argv)
{
  int acc_cp_0_size;
  int acc_cp_1_size;
  acc_cp_0_size = 64 * 1 * 4;
  acc_cp_1_size = 256;
  acc_corenum = acc_get_num_threads();
  acc_myid = acc_get_thread_num();
  acc_check_ldm(0 + sizeof(_ldm_B) + sizeof(_ldm_i) + sizeof(_ldm_j) + sizeof(acc_data_ret) + sizeof(_ldm_A));
  acc_ret = acc_sync_pe_m2l_nostride_copy(_ldm_A,A,acc_cp_0_size);
{
    for (acc_blockindex_i0_0 = acc_myid * 1; acc_blockindex_i0_0 <= 127; acc_blockindex_i0_0 += 1 * acc_corenum) {
      for (_ldm_i = 0; _ldm_i <= 0; _ldm_i += 1) {
        for (_ldm_j = 0; _ldm_j <= 63; _ldm_j += 1) {
          _ldm_B[_ldm_i][_ldm_j] = _ldm_A[_ldm_j] + (_ldm_i + acc_blockindex_i0_0);
        }
      }
      _swacc_ret = acc_sync_pe_l2m_nostride_copy(&B[acc_blockindex_i0_0][0],_ldm_B[0],acc_cp_1_size);
    }
  }
}

注意 acc_ret = acc_sync_pe_m2l_nostride_copy(_ldm_A,A,acc_cp_0_size);
这就是完全拷贝 A

loop不能和copy配合

写了一个很简单的二重循环，采用openacc加速

#define WIDTH 1024
#define HEIGHT 1024


void timer_start(struct timeval *timev);
void timer_end(double *sec,struct timeval *timev);


int main() {
   struct timeval t1_start;
   double t1;
   int y,x;
   float image[HEIGHT][WIDTH];

 // FILE *fp=fopen("image.pgm","wb");  fprintf(fp,"P5\n%s\n%d %d\n%d\n","#comment",WIDTH,HEIGHT,MAX_COLOR);

  timer_start(&t1_start);
#pragma acc parallel loop  copyout(image) 
  for(y=0;y<HEIGHT;y++) {
    for(x=0;x<WIDTH;x++) {
       image[y][x]= x+y;
    }
  }
  timer_end(&t1,&t1_start);



  printf("Time: %lf seconds.\n", t1);
  //fwrite((unsigned char*)image,sizeof(unsigned char),WIDTH*HEIGHT,fp);fclose(fp);
  return 0;
}

void timer_start(struct timeval *timev)
{
    gettimeofday(timev, 0L);
}

void timer_end(double *sec,struct timeval *timev)
{
    struct timeval t;
    gettimeofday(&t, 0L);
    *sec = 1.0*(t.tv_sec-timev->tv_sec) + 0.000001*(t.tv_usec-timev->tv_usec);
}

但是提交运行，有时候可以有时候不行

[para080@psn010 test_mandelbrot]$ bsub -I -b  -q q_sw_expr -n 1 -cgsp 64  ./a.out
Job <52832701> has been submitted to queue <q_sw_expr>
waiting for dispatch ...
dispatching ...
Time: 0.000781 seconds.
Job 52832701 has been finished.
[para080@psn010 test_mandelbrot]$ bsub -I -b  -q q_sw_expr -n 1 -cgsp 64  ./a.out
Job <52832702> has been submitted to queue <q_sw_expr>
waiting for dispatch ...
dispatching ...
Time: 0.000807 seconds.

Job 52832702 has been finished.
[para080@psn010 test_mandelbrot]$ bsub -I -b  -q q_sw_expr -n 1 -cgsp 64  ./a.out
Job <52832703> has been submitted to queue <q_sw_expr>
waiting for dispatch ...
dispatching ...
[52832703 ## Fri Jun 11 09:03:33 2021] warning: node [12]: user's mpe task: tid= 0, pid= 9119, terminated by sig 11

Job 52832703 has been finished. exit_code is 104
[para080@psn010 test_mandelbrot]$ bsub -I -b  -q q_sw_expr -n 1 -cgsp 64  ./a.out
Job <52832705> has been submitted to queue <q_sw_expr>
waiting for dispatch ...
dispatching ...
[52832705 ## Fri Jun 11 09:04:06 2021] warning: node [12]: user's mpe task: tid= 0, pid= 9133, terminated by sig 11
Job 52832705 has been finished. exit_code is 104
[para080@psn010 test_mandelbrot]$ bsub -I -b  -q q_sw_expr -n 1 -cgsp 64  ./a.out
Job <52832707> has been submitted to queue <q_sw_expr>
waiting for dispatch ...
dispatching ...
Time: 0.000779 seconds.
Job 52832707 has been finished.

u如果哦将 image 定义为 double 就肯定不行