我正在学习OPENMP并编写以下代码来解决nqueens问题。
//Full Code: https://github.com/Shafaet/Codes/blob/master/OPENMP/Parallel%20N- Queen%20problem.cpp
int n;
int call(int col,int rowmask,int dia1,int dia2)
{
if(col==n)
{
return 1;
}
int row,ans=0;
for(row=0;row<n;row++)
{
if(!(rowmask & (1<<row)) & !(dia1 & (1<<(row+col))) & !(dia2 & (1<<((row+n-1)-col))))
{
ans+=call(col+1,rowmask|1<<row,dia1|(1<<(row+col)), dia2|(1<<((row+n-1)-col)));
}
}
return ans;
}
double parallel()
{
double st=omp_get_wtime();
int ans=0;
int i;
int rowmask=0,dia1=0,dia2=0;
#pragma omp parallel for reduction(+:ans) shared(i,rowmask)
for(i=0;i<n;i++)
{
rowmask=0;
dia1=0,dia2=0;
int col=0,row=i;
ans+=call(1,rowmask|1<<row,dia1|(1<<(row+col)), dia2|(1<<((row+n-1)-col)));
}
printf("Found %d configuration for n=%d\n",ans,n);
double en=omp_get_wtime();
printf("Time taken using openmp %lf\n",en-st);
return en-st;
}
double serial()
{
double st=omp_get_wtime();
int ans=0;
int i;
int rowmask=0,dia1=0,dia2=0;
for(i=0;i<n;i++)
{
rowmask=0;
dia1=0,dia2=0;
int col=0,row=i;
ans+=call(1,rowmask|1<<row,dia1|(1<<(row+col)), dia2|(1<<((row+n-1)-col)));
}
printf("Found %d configuration for n=%d\n",ans,n);
double en=omp_get_wtime();
printf("Time taken without openmp %lf\n",en-st);
return en-st;
}
int main()
{
double average=0;
int count=0;
for(int i=2;i<=13;i++)
{
count++;
n=i;
double stime=serial();
double ptime=parallel();
printf("OpenMP is %lf times faster for n=%d\n",stime/ptime,n);
average+=stime/ptime;
puts("===============");
}
printf("On average OpenMP is %lf times faster\n",average/count);
return 0;
}
并行代码已经比普通代码更快但我想知道如何使用openmp pragma更好地优化它。我想知道我应该做些什么以获得更好的表现以及我不该做的事情。
提前致谢。
(请不要建议任何与并行编程无关的优化)
答案 0 :(得分:0)
你的代码似乎使用经典的回溯N-Queens递归算法,这对于N-Queens求解并不是最快的,但是(由于简单性)是练习方面最生动的算法并行性基础知识。 这就是说:这很简单,因此你不要指望它自然地展示了许多先进的OpenMP手段,除了基本的“并行”和减少。
但是,只要您正在寻找学习并行性,并且可能为了更清晰和更好的学习曲线,还有一个(在许多可能的实现中)可用的实现,它使用相同的算法但从教育角度来看,往往更具可读性和生动性:
void setQueen(int queens[], int row, int col) {
//check all previously placed rows for attacks
for(int i=0; i<row; i++) {
// vertical attacks
if (queens[i]==col) {
return;
}
// diagonal attacks
if (abs(queens[i]-col) == (row-i) ) {
return;
}
}
// column is ok, set the queen
queens[row]=col;
if(row==size-1) {
#pragma omp atomic
nrOfSolutions++; //Placed final queen, found a solution
}
else {
// try to fill next row
for(int i=0; i<size; i++) {
setQueen(queens, row+1, i);
}
}
}
//Function to find all solutions for nQueens problem on size x size chessboard.
void solve() {
#pragma omp parallel for
for(int i=0; i<size; i++) {
// try all positions in first row
int * queens = new int[size]; //array representing queens placed on a chess board. Index is row position, value is column.
setQueen(queens, 0, i);
delete[](queens);
}
}
这个给定的代码是Intel Advisor XE个样本之一(对于C ++和Fortran);给定样本的并行化方面将在给定Parallel Programming Book的第10章中以非常详细的方式进行讨论(事实上,给定章节仅使用N-Queens来演示如何使用工具来并行化串行代码 )。
鉴于Advisor n-queens样本使用与您的算法基本相同的算法,但它用+ atomic的简单并行组合替换了显式缩减。预计此代码效率较低,但更具“程序风格”且更具“教育性”,因为它展示了“隐藏”数据竞争。如果您上传给定的样本代码,您实际上将使用TBB,Cilk Plus和OpenMP(OMP用于C ++和Fortran)找到4个等效的N-Queens并行实现。
答案 1 :(得分:0)
我知道我对派对来说有点晚了,但你可以使用任务排队进行进一步的优化。(结果大约快7-10%)。不知道为什么。这是我正在使用的代码:
#include <iostream> // std::cout, cin, cerr ...
#include <iomanip> // modify std::out
#include <omp.h>
using namespace std;
int nrOfSolutions=0;
int size=0;
void print(int queens[]) {
cerr << "Solution " << nrOfSolutions << endl;
for(int row=0; row<size; row++) {
for(int col=0; col<size; col++) {
if(queens[row]==col) {
cout << "Q";
}
else {
cout << "-";
}
}
cout << endl;
}
}
void setQueen(int queens[], int row, int col, int id) {
for(int i=0; i<row; i++) {
// vertical attacks
if (queens[i]==col) {
return;
}
// diagonal attacks
if (abs(queens[i]-col) == (row-i) ) {
return;
}
}
// column is ok, set the queen
queens[row]=col;
if(row==size-1) {
// only one thread should print allowed to print at a time
{
// increasing the solution counter is not atomic
#pragma omp critical
nrOfSolutions++;
#ifdef _DEBUG
#pragma omp critical
print(queens);
#endif
}
}
else {
// try to fill next row
for(int i=0; i<size; i++) {
setQueen(queens, row+1, i, id);
}
}
}
void solve() {
int myid=0 ;
#pragma omp parallel
#pragma omp single
{
for(int i=0; i<size; i++) {
/*
#ifdef _OMP //(???)
myid = omp_get_thread_num();
#endif
#ifdef _DEBUG
cout << "ThreadNum: " << myid << endl ;
#endif
*/
// try all positions in first row
// create separate array for each recursion
// started here
#pragma omp task
setQueen(new int[size], 0, i, myid);
}
}
}
int main(int argc, char*argv[]) {
if(argc !=2) {
cerr << "Usage: nq-openmp-taskq boardSize.\n";
return 0;
}
size = atoi(argv[1]);
cout << "Starting OpenMP Task Queue solver for size " << size << "...\n";
double st=omp_get_wtime();
solve();
double en=omp_get_wtime();
printf("Time taken using openmp %lf\n",en-st);
cout << "Number of solutions: " << nrOfSolutions << endl;
return 0;
}