OpenACC parallel

▶ 使用 kernels 导语并行化 for 循环

● 同一段代码，使用 kernels，parallel 和 parallel + loop 进行对比

 #include <stdio.h>
 #include <time.h>
 #include <openacc.h>

 const int row = ;

 int main()
 {
     int i, j, k, a[row], b[row], c[row];
     clock_t time;
     for (i = ; i < row; i++)
         a[i] = b[i] = i;

 #ifdef _OPENACC
     time = clock();
 #pragma acc kernels     // 使用 kernels 或 parallel 或 parallel + loop
 // #pragma acc parallel
 // #pragma acc loop
     for (i = ; i < row; i++)
         c[i] = a[i] + b[i];
     time = clock() - time;
     printf("\nTime with acc:%d ms\n", time);
 #else
     time = clock();
     for (i = ; i < row; i++)
         c[i] = a[i] + b[i];
     time = clock() - time;
     printf("\nTime without acc:%d ms\n", time);
 #endif
     getchar();
     return ;
 }

● 输出结果

D:\Code\OpenACC\OpenACCProject\OpenACCProject>pgcc -acc -Minfo main.c -o main_acc_kernels.exe       // kernels
main:
     , Generating implicit copyin(b[:row])
         Generating implicit copyout(c[:row])
         Generating implicit copyin(a[:row])
     , Loop is parallelizable
         Accelerator kernel generated
         Generating Tesla code
         , #pragma acc loop gang, vector(128) /* blockIdx.x threadIdx.x */

D:\Code\OpenACC\OpenACCProject\OpenACCProject>pgcc -acc -Minfo main.c -o main_acc_parallel.exe      // parallel
main:
     , Accelerator kernel generated
         Generating Tesla code
         , #pragma acc loop vector(128) /* threadIdx.x */
     , Generating implicit copyout(c[:row])
         Generating implicit copyin(b[:row],a[:row])
     , Loop is parallelizable

D:\Code\OpenACC\OpenACCProject\OpenACCProject>pgcc -acc -Minfo main.c -o main_acc_parallel_loop.exe // parallel + loop
main:
     , Accelerator kernel generated
         Generating Tesla code
         , #pragma acc loop gang, vector(128) /* blockIdx.x threadIdx.x */
     , Generating implicit copyout(c[:row])
         Generating implicit copyin(b[:row],a[:row])

D:\Code\OpenACC\OpenACCProject\OpenACCProject>main_acc_kernels.exe
launch CUDA kernel  file=C:/Program Files (x86)/Windows Kits//Include/10.0.16299.0/ucrt\time.h function=main
line= device= threadid= num_gangs= num_workers= vector_length= grid= block=    // 多个 gang，自动配置，线程网格全都是一维的

Time with acc: ms

D:\Code\OpenACC\OpenACCProject\OpenACCProject>main_acc_parallel.exe
launch CUDA kernel  file=C:/Program Files (x86)/Windows Kits//Include/10.0.16299.0/ucrt\time.h function=main
line= device= threadid= num_gangs= num_workers= vector_length= grid= block=        // 一个 gang，gang冗余模式

Time with acc: ms

D:\Code\OpenACC\OpenACCProject\OpenACCProject>main_acc_parallel_loop.exe
launch CUDA kernel  file=C:/Program Files (x86)/Windows Kits//Include/10.0.16299.0/ucrt\time.h function=main
line= device= threadid= num_gangs= num_workers= vector_length= grid= block=    // 多个 gang，gang分裂模式

Time with acc: ms

● 二重循环，考虑是否在内层循环中使用 loop 导语

 #include <stdio.h>
 #include <time.h>
 #include <openacc.h>

 const int row = , col = ;

 int main()
 {
     int i, j, k, a[row][col], b[row][col], c[row][col];
     clock_t time;
     for (i = ; i < row; i++)
     {
         for (j = ; j < col; j++)
             a[i][j] = b[i][j] = i + j;
     }

 #ifdef _OPENACC
     time = clock();
 #pragma acc parallel
 #pragma acc loop
     for (i = ; i < row; i++)
     {
 // #pragma acc loop
         for (j = ; j < col; j++)
             c[i][j] = a[i][j] + b[i][j];
     }
     time = clock() - time;
     printf("\nTime with acc:%d ms\n", time);
 #else
     time = clock();
     for (i = ; i < row; i++)
     {
         for (j = ; j < col; j++)
             c[i][j] = a[i][j] + b[i][j];
     }
     time = clock() - time;
     printf("\nTime without acc:%d ms\n", time);
 #endif
     getchar();
     return ;
 }

● 输出结果

D:\Code\OpenACC\OpenACCProject\OpenACCProject>pgcc -acc -Minfo main.c -o main_acc_loop1.exe // 仅使用外层 loop
main:
     , Accelerator kernel generated
         Generating Tesla code
         , #pragma acc loop gang /* blockIdx.x */
         , #pragma acc loop vector(128) /* threadIdx.x */
     , Generating implicit copyin(a[:row][:col])
         Generating implicit copyout(c[:row][:col])
         Generating implicit copyin(b[:row][:col])
     , Loop is parallelizable

D:\Code\OpenACC\OpenACCProject\OpenACCProject>pgcc -acc -Minfo main.c -o main_acc_loop2.exe // 内外都使用 loop，优化结果完全相同
main:
     , Accelerator kernel generated
         Generating Tesla code
         , #pragma acc loop gang /* blockIdx.x */
         , #pragma acc loop vector(128) /* threadIdx.x */
     , Generating implicit copyin(a[:row][:col])
         Generating implicit copyout(c[:row][:col])
         Generating implicit copyin(b[:row][:col])
     , Loop is parallelizable

D:\Code\OpenACC\OpenACCProject\OpenACCProject>main_acc_loop1.exe
launch CUDA kernel  file=C:/Program Files (x86)/Windows Kits//Include/10.0.16299.0/ucrt\time.h function=main
line= device= threadid= num_gangs= num_workers= vector_length= grid= block=

Time with acc: ms

D:\Code\OpenACC\OpenACCProject\OpenACCProject>main_acc_loop2.exe
launch CUDA kernel  file=C:/Program Files (x86)/Windows Kits//Include/10.0.16299.0/ucrt\time.h function=main
line= device= threadid= num_gangs= num_workers= vector_length= grid= block=  // 优化结果完全相同

Time with acc: ms

● 三重循环，无论仅使用外循环 loop、外中循环 loop，还是外中内循环 loop，获得的编译和运行结果都是相同的，只放上来一个进行讨论

 #include <stdio.h>
 #include <time.h>
 #include <openacc.h>

 const int row = , col = , page = ;

 int main()
 {
     int i, j, k, a[row][col][page], b[row][col][page], c[row][col][page];
     clock_t time;
     for (i = ; i < row; i++)
     {
         for (j = ; j < col; j++)
         {
             for (k = ; k < page; k++)
                 a[i][j][k] = b[i][j][k] = i + j + k;
         }
     }

 #ifdef _OPENACC
     time = clock();
 #pragma acc parallel
 #pragma acc loop
     for (i = ; i < row; i++)
     {
 //#pragma acc loop
         for (j = ; j < col; j++)
         {
 //#pragma acc loop
             for (k = ; k<page; k++)
                 c[i][j][k] = a[i][j][k] + b[i][j][k];
         }
     }
     time = clock() - time;
     printf("\nTime with acc:%d ms\n", time);
 #else
     time = clock();
     for (i = ; i < row; i++)
     {
         for (j = ; j < col; j++)
         {
             for (k = ; k<page; k++)
                 c[i][j][k] = a[i][j][k] + b[i][j][k];
         }
     }
     time = clock() - time;
     printf("\nTime without acc:%d ms\n", time);
 #endif
     getchar();
     return ;
 }

● 输出结果

D:\Code\OpenACC\OpenACCProject\OpenACCProject>pgcc -acc -Minfo main.c -o main_acc_loop.exe
main:
     , Accelerator kernel generated
         Generating Tesla code
         , #pragma acc loop gang /* blockIdx.x */ // 并行化了外层循环和内层循环，但是用中间层使用的是串行
         , #pragma acc loop seq
         , #pragma acc loop vector(128) /* threadIdx.x */
     , Generating implicit copyout(c[:row][:col][:page])
         Generating implicit copyin(b[:row][:col][:page],a[:row][:col][:page])
     , Loop is parallelizable
     , Loop is parallelizable

D:\Code\OpenACC\OpenACCProject\OpenACCProject>main_acc_loop1.exe
launch CUDA kernel  file=C:/Program Files (x86)/Windows Kits//Include/10.0.16299.0/ucrt\time.h function=main
line= device= threadid= num_gangs= num_workers= vector_length= grid= block=

Time with acc: ms