OpenACC kernels

▶ 使用 kernels 导语并行化 for 循环

● 一重循环

 #include <stdio.h>
 #include <time.h>
 #include <openacc.h>
 
 const int row =  *  * ;
 
 int main()
 {
     int a[row], b[row], c[row];
     for (int i = ; i < row; ++i)                // 填充 a 和 b
         a[i] = b[i] = i;
 
     clock_t time = clock();
 #ifdef _OPENACC                                  // 使用 OpenACC 时执行本段
 #pragma acc kernels
     for (int i = ; i < row; ++i)                // c = a + b
         c[i] = a[i] + b[i];
     time = clock() - time;
     printf("\nTime with acc:%d ms\n", time);
 #else                                            // 不用 OpenACC 时执行本段
     for (int 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>pgcc main.c -o main-no-acc.exe -Minfo                                   // 编译，-Minfo 要求输出编译优化信息，没有额外输出
 
D:\Code\OpenACC>pgcc main.c -o main.exe -Minfo -acc                                     // 编译，-acc 要求使用 OpenACC
main:
     , Generating implicit copyin(b[:row])                                            // 数据管理控制
         Generating implicit copyout(c[:row])
         Generating implicit copyin(a[:row])
     , Loop is parallelizable                                                         // 并行优化
         Generating Tesla code
         , #pragma acc loop gang, vector(128) /* blockIdx.x threadIdx.x */            // 使用默认 vector 尺寸，注释是自动生成的
 
D:\Code\OpenACC>main-no-acc.exe
 
Time without acc: ms
 
D:\Code\OpenACC>main-acc.exe
launch CUDA kernel  file=D:\Code\OpenACC\main.c function=main line= device= threadid= num_gangs= num_workers= vector_length= grid= block=
                                                                                        // 对代码第 16 行的 for 进行了并行优化，
Time with acc: ms                                                                    // 使用第 0 号设备（GPU）
                                                                                        // 线程编号 1，使用 gang 65536 个，worker 1 个，vector 宽度 128
                                                                                        // CUDA 配置为 gridDim.x = 65536，blockDim.x = 128Time
                                                                                        // 每单元计算负载 = row / grid / block = 2

● 二重循环

 #include <stdio.h>
 #include <time.h>
 #include <openacc.h>
 
 const int row =  * , col = ;
 
 int main()
 {
     int a[row][col], b[row][col], c[row][col];
     for (int i = ; i < row; i++)                // 填充 a 和 b
     {
         for (int j = ; j < col; j++)
             a[i][j] = b[i][j] = i * j;
     }
 
     clock_t time = clock();
 #ifdef _OPENACC
 #pragma acc kernels
     for (int i = ; i < row; i++)                // c = a + b
     {
         for (int 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
     for (int i = ; i < row; i++)
     {
         for (int 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>pgcc main.c -o main-no-acc.exe -Minfo
 
D:\Code\OpenACC>pgcc main.c -o main-acc.exe -Minfo -acc
main:
     , Generating implicit copyin(a[:row][:col])
         Generating implicit copyout(c[:row][:col])
         Generating implicit copyin(b[:row][:col])
     , Loop is parallelizable
     , Loop is parallelizable
         Generating Tesla code
         , #pragma acc loop gang, vector(4) /* blockIdx.y threadIdx.y */      // 高一层的循环使用的是 worker
         , #pragma acc loop gang, vector(32) /* blockIdx.x threadIdx.x */
 
D:\Code\OpenACC>main-no-acc.exe
 
Time without acc: ms
 
D:\Code\OpenACC>main-acc.exe
launch CUDA kernel  file=D:\Code\OpenACC\main.c function=main line= device= threadid= num_gangs= num_workers= vector_length= grid=16x2048 block=32x4
                                                                                // 注意参数变化，仍有 num_gangs = grid，num_workers * vector_length = block
Time with acc: ms                                                            // 每单元计算负载 = row * col / grid / block = 4

● 三重循环

 #include <stdio.h>
 #include <time.h>
 #include <openacc.h>
 
 const int row = , col = , page = ;
 
 int main()
 {
     int a[row][col][page], b[row][col][page], c[row][col][page];
     for (int i = ; i < row; i++)                        // 填充 a 和 b
     {
         for (int j = ; j < col; j++)
         {
             for (int k = ; k < page; k++)
                 a[i][j][k] = b[i][j][k] = i * j + k;
         }
     }
     clock_t time = clock();
 #ifdef _OPENACC
 #pragma acc kernels
     for (int i = ; i < row; i++)                        // c = a + b
     {
         for (int j = ; j < col; j++)
         {
             for (int 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
     for (int i = ; i < row; i++)
     {
         for (int j = ; j < col; j++)
         {
             for (int 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>pgcc main.c -o main-no-acc.exe -Minfo
 
D:\Code\OpenACC>pgcc main.c -o main-acc.exe -Minfo -acc
main:
     , Generating implicit copyin(b[:row][:col][:page])
         Generating implicit copyout(c[:row][:col][:page])
         Generating implicit copyin(a[:row][:col][:page])
     , Loop is parallelizable
     , Loop is parallelizable
     , Loop is parallelizable
         Generating Tesla code
         , #pragma acc loop gang /* blockIdx.y */                             // 最高层循环尝试调整 grid
         , #pragma acc loop gang, vector(4) /* blockIdx.z threadIdx.y */
         , #pragma acc loop gang, vector(32) /* blockIdx.x threadIdx.x */
 
D:\Code\OpenACC>main-no-acc.exe
 
Time without acc: ms
 
D:\Code\OpenACC>main-acc.exe
launch CUDA kernel  file=D:\Code\OpenACC\main.c function=main line= device= threadid= num_gangs= num_workers= vector_length= grid=16x128x16 block=32x4
                                                                                // grid 变成了三维
Time with acc: ms                                                            // 每单元计算负载 = row *col * page / grid / block = 4
                                                                                // row 改为 64，则 grid=16x128x16 block=32x4，计算负载 = 2
                                                                                // col 改为 128，则 grid=16x128x16 block=32x4，计算负载 = 2
                                                                                // page 改为 256，则 grid=8x128x32 block=32x4，计算负载 = 2
                                                                                // row 改为 32，则 grid=16x32x64 block=32x4，计算负载 = 1

● 在 ubuntu 上跑一重循环的代码，注意计时器单位是 μs

cuan@CUAN:~/Temp$ pgcc -acc main.c -o main.exe
cuan@CUAN:~/Temp$ pgcc main.c -o main-no-acc.exe
cuan@CUAN:~/Temp$ ./main.exe 
 
Time with acc: us
 
cuan@CUAN:~/Temp$ ./main-no-acc.exe 
 
Time without acc: us