//?SSE_LU.cpp?:?定義控制臺應用程序的入口點。
//

#include?“stdafx.h“
#include?
#include?
#include?
#include?
#include?
#include?
#include?
#include?
#define?N?4096

using?namespace?std;

float**?serial（float?**A）?//************************************串行的高斯*******************************************
{
	for?（int?k?=?0;?k?	{
		float?tmp?=A[k][k];
		for?（int?j?=?k;?j?		{
			A[k][j]?=?A[k][j]?/?tmp;
		}
		for?（int?i?=?k?+?1;?i?		{
			float?tmp2?=?A[i][k];
			for?（int?j?=?k?+?1;?j?			{
				A[i][j]?=?A[i][j]?-?tmp2?*?A[k][j];
			}
			A[i][k]?=?0;
		}
	}
	return?A;
}

void?parallel（float?**A）?//************************************并行的高斯*****************************************
{
	__m128?t1?t2?t3?t4;
	for?（int?k?=?0;?k?	{
		float?tmp[4]?=?{?A[k][k]?A[k][k]?A[k][k]?A[k][k]?};
		t1?=?_mm_loadu_ps（tmp）;
		for?（int?j?=?N?-?4;?j?>=?k;?j?-=?4）?//從后向前每次取四個
		{
			t2?=?_mm_loadu_ps（A[k]?+?j）;
			t3?=?_mm_div_ps（t2?t1）;//除法
			_mm_storeu_ps（A[k]?+?j?t3）;
		}
		if?（k?%?4?!=?（N?%?4））?//處理不能被4整除的元素
		{
			for?（int?j?=?k;?j?%?4?!=?（?N%?4）;?j++）
			{
				A[k][j]?=?A[k][j]?/?tmp[0];
			}
		}
		for?（int?j?=?（N?%?4）?-?1;?j?>=?0;?j--）?
		{????
			A[k][j]?=?A[k][j]?/?tmp[0];
		}
		for?（int?i?=?k?+?1;?i?		{
			float?tmp[4]?=?{?A[i][k]?A[i][k]?A[i][k]?A[i][k]?};
			t1?=?_mm_loadu_ps（tmp）;
			for?（int?j?=?N?-?4;?j?>k;?j?-=?4）
			{
				t2?=?_mm_loadu_ps（A[i]?+?j）;
				t3?=?_mm_loadu_ps（A[k]?+?j）;
				t4?=?_mm_sub_ps（t2?_mm_mul_ps（t1?t3））;?//減法
				_mm_storeu_ps（A[i]?+?j?t4）;
			}
			for?（int?j?=?k?+?1;?j?%?4?!=?（N?%?4）;?j++）
			{
				A[i][j]?=?A[i][j]?-?A[i][k]?*?A[k][j];
			}
			A[i][k]?=?0;
		}
	}
}

void?print（float?**Aint?num）?//**********************************打印矩陣*****************************************
{
	for?（int?i?=?0;?i?	{
		for?（int?j?=?0;?j?		{
			cout?<		}
		cout?<	}
}

int?_tmain（）?//*************************************主函數入口********************************************
{
	srand（（unsigned）time（NULL））;?//生成種子?
	float?**A?=?new?float*[N];
	float?**A2?=?new?float*[N];
	for?（int?i?=?0;?i	{
		A[i]?=?new?float[N];
		A2[i]?=?A[i];
	}

	cout?<	for?（int?i?=?0;?i?	{
		for?（int?j?=?0;?j?		{
			A[i][j]?=?rand（）?%?90;?//取值為1-90之間
		}
	}
	print（A10）;

	cout?<	clock_t??clockBegin?clockEnd;
	clockBegin?=?clock（）;?//開始計時
	float?**B?=?serial（A）;
	clockEnd?=?clock（）;
	print（A10）;
	cout?<
	cout?<	clockBegin?=?clock（）;?//開始計時
	parallel（A2）;
	clockEnd?=?clock（）;
	print（A210）;
	cout?<	return?

?屬性????????????大小?????日期????時間???名稱
-----------?---------??----------?-----??----

?????文件?????156783??2016-07-19?14:34??高斯消去法SSE并行化.docx

?????文件??????74752??2016-03-29?17:25??SSE_LU\Debug\SSE_LU.exe

?????文件?????708952??2016-03-29?17:25??SSE_LU\Debug\SSE_LU.ilk

?????文件????1298432??2016-03-29?17:25??SSE_LU\Debug\SSE_LU.pdb

?????文件????3080192??2016-03-29?17:25??SSE_LU\ipch\sse_lu-5967a1ae\sse_lu-b3fb2d97.ipch

?????文件???????1735??2016-03-29?17:25??SSE_LU\SSE_LU\Debug\SSE_LU.log

?????文件?????165993??2016-03-29?17:25??SSE_LU\SSE_LU\Debug\SSE_LU.obj

?????文件????1703936??2016-03-29?15:11??SSE_LU\SSE_LU\Debug\SSE_LU.pch

?????文件???????1374??2016-03-29?17:25??SSE_LU\SSE_LU\Debug\SSE_LU.tlog\cl.command.1.tlog

?????文件??????30640??2016-03-29?17:25??SSE_LU\SSE_LU\Debug\SSE_LU.tlog\CL.read.1.tlog

?????文件????????910??2016-03-29?17:25??SSE_LU\SSE_LU\Debug\SSE_LU.tlog\CL.write.1.tlog

?????文件???????1208??2016-03-29?17:25??SSE_LU\SSE_LU\Debug\SSE_LU.tlog\link.command.1.tlog

?????文件???????2738??2016-03-29?17:25??SSE_LU\SSE_LU\Debug\SSE_LU.tlog\link.read.1.tlog

?????文件????????480??2016-03-29?17:25??SSE_LU\SSE_LU\Debug\SSE_LU.tlog\link.write.1.tlog

?????文件????????188??2016-03-29?17:25??SSE_LU\SSE_LU\Debug\SSE_LU.tlog\SSE_LU.lastbuildstate

?????文件??????11920??2016-03-29?15:11??SSE_LU\SSE_LU\Debug\stdafx.obj

?????文件?????764928??2016-03-29?17:25??SSE_LU\SSE_LU\Debug\vc120.idb

?????文件?????487424??2016-03-29?17:25??SSE_LU\SSE_LU\Debug\vc120.pdb

?????文件???????1503??2016-03-29?14:05??SSE_LU\SSE_LU\ReadMe.txt

?????文件???????3110??2016-03-29?17:25??SSE_LU\SSE_LU\SSE_LU.cpp

?????文件???????4531??2016-03-29?14:05??SSE_LU\SSE_LU\SSE_LU.vcxproj

?????文件???????1313??2016-03-29?14:05??SSE_LU\SSE_LU\SSE_LU.vcxproj.filters

?????文件????????212??2016-03-29?14:05??SSE_LU\SSE_LU\stdafx.cpp

?????文件????????234??2016-03-29?14:05??SSE_LU\SSE_LU\stdafx.h

?????文件????????236??2016-03-29?14:05??SSE_LU\SSE_LU\targetver.h

?????文件???33619968??2016-03-29?17:34??SSE_LU\SSE_LU.sdf

?????文件????????964??2016-03-29?14:05??SSE_LU\SSE_LU.sln

????..A..H.?????13824??2016-03-29?17:34??SSE_LU\SSE_LU.v12.suo

?????目錄??????????0??2016-03-29?17:25??SSE_LU\SSE_LU\Debug\SSE_LU.tlog

?????目錄??????????0??2016-03-29?14:05??SSE_LU\ipch\sse_lu-5967a1ae

............此處省略8個文件信息