OpenACC 优化矩阵乘法
▶ 按书上的步骤使用不同的导语优化矩阵乘法
● 已经优化的代码
#include <iostream>
#include <cstdlib>
#include <chrono> #define SIZE 1024 using namespace std;
using namespace std::chrono; double a[SIZE][SIZE], b[SIZE][SIZE], c[SIZE][SIZE], d[SIZE][SIZE]; // 四个数组放入 main 里会报错 Segmentation fault (core dumped) int main()
{
//int i, j, k; // ijk 和 tmp 在循环中使用时才声明会导致运行时间变长
double tmp; #pragma acc enter data create(a, b, c)
#pragma acc kernels present(a, b, c)
{
for (int i = ; i < SIZE; i++) // 初始化 ab
{
for (int j = ; j < SIZE; j++)
a[i][j] = (double)(i + j);
}
for (int i = ; i < SIZE; i++) // 初始化 ab
{
for (int j = ; j < SIZE; j++)
b[i][j] = (double)(i - j);
}
for (int i = ; i < SIZE; i++) // 每种方法前都要清空 c
{
for (int j = ; j < SIZE; j++)
c[i][j] = 0.0;
}
} high_resolution_clock::time_point t1 = high_resolution_clock::now(); #pragma acc kernels present(a, b, c) // 方法 1，每层循环都 auto
{
#pragma acc loop auto
for (int i = ; i < SIZE; i++)
{
#pragma acc loop auto
for (int j = ; j < SIZE; j++)
{
#pragma acc loop auto
for (int k = ; k < SIZE; k++)
c[i][j] += a[i][k] * b[k][j];
}
}
} high_resolution_clock::time_point t2 = high_resolution_clock::now();
duration<double> time = duration_cast<duration<double>>(t2 - t1);
printf("Time OpenACC - Auto: %.6lf s.\n\n", time.count()); #pragma acc kernels present(c)
for (int i = ; i < SIZE; i++)
{
for (int j = ; j < SIZE; j++)
c[i][j] = 0.0;
} t1 = high_resolution_clock::now(); #pragma acc kernels present(a, b, c) // 方法 2，外两层 independent，最里层串行
{
#pragma acc loop independent
for (int i = ; i < SIZE; i++)
{
#pragma acc loop independent
for (int j = ; j < SIZE; j++)
{
#pragma acc loop independent
for (int k = ; k < SIZE; k++)
c[i][j] += a[i][k] * b[k][j];
}
}
} t2 = high_resolution_clock::now();
time = duration_cast<duration<double>>(t2 - t1);
printf("Time OpenACC - Independent Seq: %.6lf s.\n\n", time.count()); #pragma acc kernels present(c)
for (int i = ; i < SIZE; i++)
{
for (int j = ; j < SIZE; j++)
c[i][j] = 0.0;
} t1 = high_resolution_clock::now(); #pragma acc kernels present(a, b, c) // 方法 3，外两层 independent，最里层规约
{
#pragma acc loop independent
for (int i = ; i < SIZE; i++)
{
#pragma acc loop independent
for (int j = ; j < SIZE; j++)
{
tmp = 0.0f;
#pragma acc loop reduction(+: tmp)
for (int k = ; k < SIZE; k++)
tmp += a[i][k] * b[k][j];
c[i][j] = tmp;
}
}
} t2 = high_resolution_clock::now();
time = duration_cast<duration<double>>(t2 - t1);
printf("Time OpenACC - Independent Reduction: %.6lf s.\n\n", time.count()); #pragma acc kernels present(c)
for (int i = ; i < SIZE; i++)
{
for (int j = ; j < SIZE; j++)
c[i][j] = 0.0;
} t1 = high_resolution_clock::now(); #pragma acc kernels present(a, b, c) // 方法 4，手动指定 gang 和 vector
{
#pragma acc loop gang(32)
for (int i = ; i < SIZE; i++)
{
#pragma acc loop vector(16)
for (int j = ; j < SIZE; j++)
{
tmp = 0.0f;
#pragma acc loop reduction(+: tmp)
for (int k = ; k < SIZE; k++)
tmp += a[i][k] * b[k][j];
c[i][j] = tmp;
}
}
} t2 = high_resolution_clock::now();
time = duration_cast<duration<double>>(t2 - t1);
printf("Time OpenACC - Gang Vector: %.6lf s.\n\n", time.count()); #pragma acc kernels present(c)
for (int i = ; i < SIZE; i++)
{
for (int j = ; j < SIZE; j++)
c[i][j] = 0.0;
} t1 = high_resolution_clock::now(); #pragma acc kernels present(a, b, c) // 方法 5，分块重排
{
#pragma acc loop tile(32, 32)
for (int i = ; i < SIZE; i++)
{
for (int j = ; j < SIZE; j++)
{
tmp = 0.0f;
#pragma acc loop reduction(+ \
: tmp)
for (int k = ; k < SIZE; ++k)
tmp += a[i][k] * b[k][j];
c[i][j] = tmp;
}
}
} t2 = high_resolution_clock::now();
time = duration_cast<duration<double>>(t2 - t1);
printf("Time OpenACC - tile: %.6lf s.\n\n", time.count()); #pragma acc kernels present(c)
for (int i = ; i < SIZE; i++)
{
for (int j = ; j < SIZE; j++)
c[i][j] = 0.0;
} t1 = high_resolution_clock::now(); #pragma acc kernels present(a, b, c) // 方法 6，合并多层迭代
{
#pragma acc loop collapse(2) independent
for (int i = ; i < SIZE; i++)
{
for (int j = ; j < SIZE; j++)
{
tmp = 0.0f;
#pragma acc loop reduction(+: tmp)
for (int k = ; k < SIZE; k++)
tmp += a[i][k] * b[k][j];
c[i][j] = tmp;
}
}
} t2 = high_resolution_clock::now();
time = duration_cast<duration<double>>(t2 - t1);
printf("Time OpenACC - Collapse: %.6lf s.\n\n", time.count()); #pragma acc exit data copyout(a, b, c) #pragma omp parallel for shared(d)
for (int i = ; i < SIZE; i++)
{
for (int j = ; j < SIZE; j++)
d[i][j] = 0.0;
} t1 = high_resolution_clock::now(); #pragma omp parallel for default(none) shared(a, b, d) // 使用 OpenMP
for (int i = ; i < SIZE; i++)
{
for (int j = ; j < SIZE; j++)
{
for (int k = ; k < SIZE; k++)
d[i][j] += a[i][k] * b[k][j];
}
}
t2 = high_resolution_clock::now();
time = duration_cast<duration<double>>(t2 - t1);
printf("Time OpenMP: %.6lf s.\n\n", time.count()); for (int i = ; i < SIZE; i++) // 检查结果
{
for (int j = ; j < SIZE; j++)
{
if (c[i][j] != d[i][j])
printf("\nError at [%d, %d],c = %f d = %f \n", i, j, c[i][j], d[i][j]);
}
}
return ;
}
● 输出结果（数据管理优化前）
cuan@CUAN:/mnt/d/Code/OpenACC/OpenACCProject/OpenACCProject$ ./acc.exe Time OpenACC - Auto: 4.589736 s. Time OpenACC - Independent Seq: 4.823721 s. Time OpenACC - Independent Reduction: 3.669336 s. Time OpenACC - Gang Vector: 3.611391 s. Time OpenACC - tile: 3.609573 s. Time OpenACC - Collapse: 3.605792 s. Time OpenMP: 4.345018 s.
● 输出结果（数据管理优化后）
cuan@CUAN:~/acc$ pgc++ main.cpp -std=c++ -acc -mp -Minfo -o main.exe
main:
, include "chrono"
, include "chrono"
, Parallel region activated
, Parallel region terminated
, Parallel region activated
, Parallel region terminated
, Generating enter data create(b[:][:],c[:][:],a[:][:])
Generating present(a[:][:],b[:][:],c[:][:])
, Loop is parallelizable
, Loop is parallelizable
Generating Tesla code
, #pragma acc loop gang, vector(4) /* blockIdx.y threadIdx.y */
, #pragma acc loop gang, vector(32) /* blockIdx.x threadIdx.x */
, Loop is parallelizable
, Loop is parallelizable
Generating Tesla code
, #pragma acc loop gang, vector(4) /* blockIdx.y threadIdx.y */
, #pragma acc loop gang, vector(32) /* blockIdx.x threadIdx.x */
, Loop is parallelizable
, Loop is parallelizable
Generating Tesla code
, #pragma acc loop gang, vector(4) /* blockIdx.y threadIdx.y */
, #pragma acc loop gang, vector(32) /* blockIdx.x threadIdx.x */
, Memory zero idiom, loop replaced by call to __c_mzero8
, Generating present(a[:][:],c[:][:],b[:][:])
, Loop is parallelizable
, Loop is parallelizable
, Complex loop carried dependence of c prevents parallelization
Loop carried dependence of c prevents parallelization
Loop carried backward dependence of c prevents vectorization
Inner sequential loop scheduled on accelerator
Generating Tesla code
, #pragma acc loop gang /* blockIdx.y */
, #pragma acc loop gang, vector(128) /* blockIdx.x threadIdx.x */
, #pragma acc loop seq
, Complex loop carried dependence of c prevents parallelization
Loop carried backward dependence of c prevents vectorization
, Generating present(c[:][:])
, Loop is parallelizable
, Loop is parallelizable
Generating Tesla code
, #pragma acc loop gang, vector(4) /* blockIdx.y threadIdx.y */
, #pragma acc loop gang, vector(32) /* blockIdx.x threadIdx.x */
, Memory zero idiom, loop replaced by call to __c_mzero8
, Generating present(a[:][:],c[:][:],b[:][:])
, Loop is parallelizable
, Loop is parallelizable
, Loop is parallelizable
Generating Tesla code
, #pragma acc loop gang /* blockIdx.z */
, #pragma acc loop gang, vector(128) /* blockIdx.x threadIdx.x */
, #pragma acc loop gang /* blockIdx.y */
, Generating present(c[:][:])
, Loop is parallelizable
, Loop is parallelizable
Generating Tesla code
, #pragma acc loop gang, vector(4) /* blockIdx.y threadIdx.y */
, #pragma acc loop gang, vector(32) /* blockIdx.x threadIdx.x */
, Memory zero idiom, loop replaced by call to __c_mzero8
, Generating present(a[:][:],c[:][:],b[:][:])
, Loop is parallelizable
, Loop is parallelizable
Generating Tesla code
, #pragma acc loop gang, vector(4) /* blockIdx.y threadIdx.y */
, #pragma acc loop gang, vector(32) /* blockIdx.x threadIdx.x */
, #pragma acc loop seq
, FMA (fused multiply-add) instruction(s) generated
, Loop is parallelizable
, Generating present(c[:][:])
, Loop is parallelizable
, Loop is parallelizable
Generating Tesla code
, #pragma acc loop gang, vector(4) /* blockIdx.y threadIdx.y */
, #pragma acc loop gang, vector(32) /* blockIdx.x threadIdx.x */
, Memory zero idiom, loop replaced by call to __c_mzero8
, Generating present(a[:][:],c[:][:],b[:][:])
, Loop is parallelizable
, Loop is parallelizable
Generating Tesla code
, #pragma acc loop gang(32), vector(8) /* blockIdx.y threadIdx.y */
, #pragma acc loop gang, vector(16) /* blockIdx.x threadIdx.x */
, #pragma acc loop seq
, Loop is parallelizable
, Generating present(c[:][:])
, Loop is parallelizable
, Loop is parallelizable
Generating Tesla code
, #pragma acc loop gang, vector(4) /* blockIdx.y threadIdx.y */
, #pragma acc loop gang, vector(32) /* blockIdx.x threadIdx.x */
, Memory zero idiom, loop replaced by call to __c_mzero8
, Generating present(a[:][:],c[:][:],b[:][:])
, Loop is parallelizable
, Loop is parallelizable
Generating Tesla code
, #pragma acc loop gang, vector tile(32,32) /* blockIdx.x threadIdx.x */
, /* blockIdx.x threadIdx.x tiled */
, #pragma acc loop seq
, Loop is parallelizable
, Generating present(c[:][:])
, Loop is parallelizable
, Loop is parallelizable
Generating Tesla code
, #pragma acc loop gang, vector(4) /* blockIdx.y threadIdx.y */
, #pragma acc loop gang, vector(32) /* blockIdx.x threadIdx.x */
, Memory zero idiom, loop replaced by call to __c_mzero8
, Generating present(a[:][:],c[:][:],b[:][:])
, Loop is parallelizable
, Loop is parallelizable
Generating Tesla code
, #pragma acc loop gang, vector(128) collapse(2) /* blockIdx.x threadIdx.x */
, /* blockIdx.x threadIdx.x collapsed */
, #pragma acc loop seq
, Loop is parallelizable
, Generating exit data copyout(c[:][:],b[:][:],a[:][:])
Parallel loop activated with static block schedule
, Memory zero idiom, loop replaced by call to __c_mzero8
, Barrier
, Parallel loop activated with static block schedule
FMA (fused multiply-add) instruction(s) generated
, Barrier
cuan@CUAN:~/acc$ ./main.exe
launch CUDA kernel file=/home/cuan/acc/main.cpp function=main line= device= threadid= num_gangs= num_workers= vector_length= grid=32x256 block=32x4
launch CUDA kernel file=/home/cuan/acc/main.cpp function=main line= device= threadid= num_gangs= num_workers= vector_length= grid=32x256 block=32x4
launch CUDA kernel file=/home/cuan/acc/main.cpp function=main line= device= threadid= num_gangs= num_workers= vector_length= grid=32x256 block=32x4
launch CUDA kernel file=/home/cuan/acc/main.cpp function=main line= device= threadid= num_gangs= num_workers= vector_length= grid=8x1024 block=
Time OpenACC - Auto: 0.018726 s. launch CUDA kernel file=/home/cuan/acc/main.cpp function=main line= device= threadid= num_gangs= num_workers= vector_length= grid=32x256 block=32x4
launch CUDA kernel file=/home/cuan/acc/main.cpp function=main line= device= threadid= num_gangs= num_workers= vector_length= grid=8x1024x4 block=
Time OpenACC - Independent Seq: 0.040719 s. launch CUDA kernel file=/home/cuan/acc/main.cpp function=main line= device= threadid= num_gangs= num_workers= vector_length= grid=32x256 block=32x4
launch CUDA kernel file=/home/cuan/acc/main.cpp function=main line= device= threadid= num_gangs= num_workers= vector_length= grid=32x256 block=32x4
Time OpenACC - Independent Reduction: 0.012491 s. launch CUDA kernel file=/home/cuan/acc/main.cpp function=main line= device= threadid= num_gangs= num_workers= vector_length= grid=32x256 block=32x4
launch CUDA kernel file=/home/cuan/acc/main.cpp function=main line= device= threadid= num_gangs= num_workers= vector_length= grid=64x32 block=16x8
Time OpenACC - Gang Vector: 0.012314 s. launch CUDA kernel file=/home/cuan/acc/main.cpp function=main line= device= threadid= num_gangs= num_workers= vector_length= grid=32x256 block=32x4
launch CUDA kernel file=/home/cuan/acc/main.cpp function=main line= device= threadid= num_gangs= num_workers= vector_length= grid= block=
Time OpenACC - tile: 0.013609 s. launch CUDA kernel file=/home/cuan/acc/main.cpp function=main line= device= threadid= num_gangs= num_workers= vector_length= grid=32x256 block=32x4
launch CUDA kernel file=/home/cuan/acc/main.cpp function=main line= device= threadid= num_gangs= num_workers= vector_length= grid= block=
Time OpenACC - Collapse: 0.012676 s. Time OpenMP: 0.504436 s. Accelerator Kernel Timing data
/home/cuan/acc/main.cpp
main NVIDIA devicenum=
time(us): ,
: compute region reached time
: kernel launched time
grid: [32x256] block: [32x4]
device time(us): total= max= min= avg=
elapsed time(us): total= max= min= avg=
: kernel launched time
grid: [32x256] block: [32x4]
device time(us): total= max= min= avg=
elapsed time(us): total= max= min= avg=
: kernel launched time
grid: [32x256] block: [32x4]
device time(us): total= max= min= avg=
elapsed time(us): total= max= min= avg=
: data region reached times
: compute region reached time
: kernel launched time
grid: [8x1024] block: []
device time(us): total=, max=, min=, avg=,
elapsed time(us): total=, max=, min=, avg=,
: data region reached times
: compute region reached time
: kernel launched time
grid: [32x256] block: [32x4]
device time(us): total= max= min= avg=
elapsed time(us): total= max= min= avg=
: data region reached times
: compute region reached time
: kernel launched time
grid: [8x1024x4] block: []
device time(us): total=, max=, min=, avg=,
elapsed time(us): total=, max=, min=, avg=,
: data region reached times
: compute region reached time
: kernel launched time
grid: [32x256] block: [32x4]
device time(us): total= max= min= avg=
elapsed time(us): total= max= min= avg=
: data region reached times
: compute region reached time
: kernel launched time
grid: [32x256] block: [32x4]
device time(us): total=, max=, min=, avg=,
elapsed time(us): total=, max=, min=, avg=,
: data region reached times
: compute region reached time
: kernel launched time
grid: [32x256] block: [32x4]
device time(us): total= max= min= avg=
elapsed time(us): total= max= min= avg=
: data region reached times
: compute region reached time
: kernel launched time
grid: [64x32] block: [16x8]
device time(us): total=, max=, min=, avg=,
elapsed time(us): total=, max=, min=, avg=,
: data region reached times
: compute region reached time
: kernel launched time
grid: [32x256] block: [32x4]
device time(us): total= max= min= avg=
elapsed time(us): total= max= min= avg=
: data region reached times
: compute region reached time
: kernel launched time
grid: [] block: []
device time(us): total=, max=, min=, avg=,
elapsed time(us): total=, max=, min=, avg=,
: data region reached times
: compute region reached time
: kernel launched time
grid: [32x256] block: [32x4]
device time(us): total= max= min= avg=
elapsed time(us): total= max= min= avg=
: data region reached times
: compute region reached time
: kernel launched time
grid: [] block: []
device time(us): total=, max=, min=, avg=,
elapsed time(us): total=, max=, min=, avg=,
: data region reached times
: data region reached time
: data copyout transfers:
device time(us): total=, max= min= avg=
OpenACC 优化矩阵乘法的更多相关文章
- Strassen优化矩阵乘法(复杂度O(n^lg7))
按照算法导论写的 还没有测试复杂度到底怎么样 不过这个真的很卡内存,挖个坑,以后写空间优化 还有Matthew Anderson, Siddharth Barman写了一个关于矩阵乘法的论文 < ...
- poj3613：Cow Relays（倍增优化+矩阵乘法floyd+快速幂）
Cow Relays Time Limit: 1000MS Memory Limit: 65536K Total Submissions: 7825 Accepted: 3068 Descri ...
- 利用Cayley-Hamilton theorem 优化矩阵线性递推
平时有关线性递推的题,很多都可以利用矩阵乘法来解决. 时间复杂度一般是O(K3logn)因此对矩阵的规模限制比较大. 下面介绍一种利用利用Cayley-Hamilton theorem加速矩阵乘法的方 ...
- [BZOJ 1009] [HNOI2008] GT考试 【AC自动机 + 矩阵乘法优化DP】
题目链接:BZOJ - 1009 题目分析 题目要求求出不包含给定字符串的长度为 n 的字符串的数量. 既然这样,应该就是 KMP + DP ,用 f[i][j] 表示长度为 i ,匹配到模式串第 j ...
- bzoj 3240: [Noi2013]矩阵游戏 矩阵乘法+十进制快速幂+常数优化
3240: [Noi2013]矩阵游戏 Time Limit: 10 Sec Memory Limit: 256 MBSubmit: 613 Solved: 256[Submit][Status] ...
- HDU 4914 Linear recursive sequence(矩阵乘法递推的优化)
题解见X姐的论文 矩阵乘法递推的优化.仅仅是mark一下. .
- [转]OpenBLAS项目与矩阵乘法优化
课程内容 OpenBLAS项目介绍 矩阵乘法优化算法 一步步调优实现 以下为公开课完整视频,共64分钟: 以下为公开课内容的文字及 PPT 整理. 雷锋网的朋友们大家好,我是张先轶,今天主要介绍一下我 ...
- 矩阵乘法优化DP
本文讲一下一些基本的矩阵优化DP的方法技巧. 定义三个矩阵A,B,C,其中行和列分别为$m\times n,n \times p,m\times p$,(其中行是从上往下数的,列是从左往右数的) $C ...
- 【BZOJ 3326】[Scoi2013]数数 数位dp+矩阵乘法优化
挺好的数位dp……先说一下我个人的做法:经过观察,发现这题按照以往的思路从后往前递增,不怎么好推,然后我就大胆猜想,从前往后推,发现很好推啊,维护四个变量,从开始位置到现在有了i个数 f[i]:所有数 ...
随机推荐
- java中 static,final,transient,volatile关键字的作用
static 和final static 静态修饰关键字,可以修饰 变量,程序块,类的方法: 当你定义一个static的变量的时候jvm会将将其分配在内存堆上,所有程序对它的引用都会指向这一个地址而 ...
- Python之进程（multiprocessing）
一.multiprocessing模块简介——进程基于“threading”的接口 multiprocessing模块支持创建进程——使用threading模块相似的API.multiprocessi ...
- LG4454 【[CQOI2018]破解D-H协议】
先谈一下BSGS算法(传送门) 但是上面这位的程序实现比较繁琐,看下面这位的. clover_hxy这样说 bsgs算法,又称大小步算法(某大神称拔山盖世算法). 主要用来解决 A^x=B(mod C ...
- MySQL Disk--SSD磁盘性能抖动问题
============================================================= SSD性能 空盘性能:SSD出厂时磁盘没有任何数据情况下的性能 稳态性能:当 ...
- STORJ 有实际应用
STORJ 有实际应用,Filezilla 支持 STORJ 的分布式协议.
- CDN、浏览器缓存
CDN是什么? 谈到CDN的作用,可以用8年买火车票的经历来形象比喻: 8年前,还没有火车票代售点一说,12306.cn更是无从说起.那时候火车票还只能在火车站的售票大厅购买,而我所住的小县城并不通火 ...
- linq to sql （Group By/Having/Count/Sum/Min/Max/Avg操作符） （转帖）
http://wenku.baidu.com/link?url=2RsCun4Mum1SLbh-LHYZpTmGFMiEukrWAoJGKGpkiHKHeafJcx2y-HVttNMb1BqJpNdw ...
- 【selenium】XPATH语法（一）
XPath 使用路径表达式来选取 XML 文档中的节点或节点集.节点是通过沿着路径 (path) 或者步 (steps) 来选取的. 我们将在下面的例子中使用这个 XML 文档. <?xml v ...
- win7 配置微软的深度学习caffe
win7 配置微软的深度学习caffe 官方下载: https://github.com/Microsoft/caffe 然后 直接修改caffe目录下的windows目录下的项目的props文件 ...
- JavaScript模块化-RequireJs实现AMD规范的简单例子
AMD规范简介 AMD(异步模块定义),是实现JavaScript模块化规范之一,它采用异步方式加载模块,模块的加载不影响后面语句的运行.require.js和curl.js都是实现AMD规范的优秀加 ...