OpenACC 计算构建内的自定义函数

▶ 使用 routine 构件创建的自定义函数，在并行调用上的差别

● 代码，自定义一个 sqab 函数，使用内建函数 fabsf 和 sqrtf 计算一个矩阵所有元素绝对值的平方根

 #include <stdio.h>
 #include <stdlib.h>
 #include <math.h>
 #include <openacc.h>

 #define ROW 8
 #define COL 64

 #pragma acc routine vector
 void sqab(float *a, const int m)
 {
 #pragma acc loop
     for (int idx = ; idx < m; idx++)
         a[idx] = sqrtf(fabsf(a[idx]));
 }

 int main()
 {
     float x[ROW][COL];
     int row, col;
     for (row = ; row < ROW; row++)
     {
         for (col = ; col < COL; col++)
             x[row][col] = row *  + col;
     }
     printf("\nx[1][1] = %f\n", x[][]);

 #pragma acc parallel loop vector pcopy(x[0:ROW][0:COL]) // 之后在这里分别添加 gang，worker，vector
     for (row = ; row < ROW; row++)
         sqab(&x[row][], COL);
     printf("\nx[1][1] = %f\n", x[][]);

     //getchar();
     return ;
 }

● 输出结果，第 28 行不添加并行级别子句（默认使用 gang）

 D:\Code\OpenACC\OpenACCProject\OpenACCProject>pgcc main.c -acc -Minfo -o main_acc.exe
 sqab:
      , Generating Tesla code
          , #pragma acc loop vector /* threadIdx.x */
      , Loop is parallelizable
 main:
      , Generating copy(x[:][:])
          Accelerator kernel generated
          Generating Tesla code
          , #pragma acc loop gang /* blockIdx.x */

 D:\Code\OpenACC\OpenACCProject\OpenACCProject>main_acc.exe

 x[][] = 11.000000
 launch CUDA kernel  file=D:\Code\OpenACC\OpenACCProject\OpenACCProject\main.c function=main
 line= device= threadid= num_gangs= num_workers= vector_length= grid= block=      // 8 个 gang 在 blockIdx.x 层级，1 个 worker，vector 在 threadIdx.x 层级

 x[][] = 3.316625
 PGI: "acc_shutdown" not detected, performance results might be incomplete.
  Please add the call "acc_shutdown(acc_device_nvidia)" to the end of your application to ensure that the performance results are complete.

 Accelerator Kernel Timing data
 D:\Code\OpenACC\OpenACCProject\OpenACCProject\main.c
   main  NVIDIA  devicenum=
     time(us):
     : compute region reached  time
         : kernel launched  time
             grid: []  block: []
             elapsed time(us): total= max= min= avg=
     : data region reached  times
         : data copyin transfers:
              device time(us): total= max= min= avg=
         : data copyout transfers:
              device time(us): total= max= min= avg=

● 输出结果，第 28 行添加并行级别子句 worker

 D:\Code\OpenACC\OpenACCProject\OpenACCProject>pgcc main.c -acc -Minfo -o main_acc.exe
 sqab:
      , Generating Tesla code
          , #pragma acc loop vector /* threadIdx.x */
      , Loop is parallelizable
 main:
      , Generating copy(x[:][:])
          Accelerator kernel generated
          Generating Tesla code
          , #pragma acc loop worker(4) /* threadIdx.y */
      , Loop is parallelizable

 D:\Code\OpenACC\OpenACCProject\OpenACCProject>main_acc.exe

 x[][] = 11.000000
 launch CUDA kernel  file=D:\Code\OpenACC\OpenACCProject\OpenACCProject\main.c function=main
 line= device= threadid= num_gangs= num_workers= vector_length= grid= block=32x4    // 1 个 gang，4 个 worker 在 threadIdx.y 层级，使用 2 维线程网格

 x[][] = 3.316625
 PGI: "acc_shutdown" not detected, performance results might be incomplete.
  Please add the call "acc_shutdown(acc_device_nvidia)" to the end of your application to ensure that the performance results are complete.

 Accelerator Kernel Timing data
 D:\Code\OpenACC\OpenACCProject\OpenACCProject\main.c
   main  NVIDIA  devicenum=
     time(us):
     : compute region reached  time
         : kernel launched  time
             grid: []  block: [32x4]
              device time(us): total= max= min= avg=
     : data region reached  times
         : data copyin transfers:
              device time(us): total= max= min= avg=
         : data copyout transfers:
              device time(us): total= max= min= avg=

● 输出结果，第 28 行添加并行级别子句 vector

 D:\Code\OpenACC\OpenACCProject\OpenACCProject>pgcc main.c -acc -Minfo -o main_acc.exe
 sqab:
      , Generating Tesla code
          , #pragma acc loop vector /* threadIdx.x */
      , Loop is parallelizable
 main:
      , Generating copy(x[:][:])
          Accelerator kernel generated
          Generating Tesla code
          , #pragma acc loop seq
      , Loop is parallelizable

 D:\Code\OpenACC\OpenACCProject\OpenACCProject>main_acc.exe

 x[][] = 11.000000
 launch CUDA kernel  file=D:\Code\OpenACC\OpenACCProject\OpenACCProject\main.c function=main
 line= device= threadid= num_gangs= num_workers= vector_length= grid= block=      // 1 个 gang，1 个 worker，并行全都堆在 threadIdx.x 层级上

 x[][] = 3.316625
 PGI: "acc_shutdown" not detected, performance results might be incomplete.
  Please add the call "acc_shutdown(acc_device_nvidia)" to the end of your application to ensure that the performance results are complete.

 Accelerator Kernel Timing data
 D:\Code\OpenACC\OpenACCProject\OpenACCProject\main.c
   main  NVIDIA  devicenum=
     time(us):
     : compute region reached  time
         : kernel launched  time
             grid: []  block: []
             elapsed time(us): total= max= min= avg=
     : data region reached  times
         : data copyin transfers:
              device time(us): total= max= min= avg=
         : data copyout transfers:
              device time(us): total= max= min= avg=

● 如果自定义函数并行子句等级高于主调函数，则主调函数并行子句会变成 seq；如果自定义函数并行子句等级低于内部并行子句等级，则会报 warning，忽略掉内部并行子句：

 #pragma acc routine vector
 void sqab(float *a, const int m)
 {
 #pragma acc loop worker
     for (int idx = ; idx < m; idx++)
         a[idx] = sqrtf(fabsf(a[idx]));
 }

● 编译结果（运行结果通上面的 worker，不写）

D:\Code\OpenACC\OpenACCProject\OpenACCProject>pgcc main.c -acc -Minfo -o main_acc.exe
PGC-W--acc loop worker clause ignored in acc routine vector procedure  (main.c: )
sqab:
     , Generating Tesla code
         , #pragma acc loop vector /* threadIdx.x */
     , Loop is parallelizable