SSE(Streaming SIMD Extensions)是英特爾在AMD的3D Now!釋出一年之後,在其計算機晶片Pentium III中引入的指令集,是MMX的超集。AMD後來在Athlon XP中加入了對這個指令集的支援。這個指令集增加了對8個128位寄存器XMM0-XMM7的支援,每個寄存器可以存儲4個單精度浮點數。使用這些寄存器的程式必須使用FXSAVE和FXRSTR指令來保持和恢複狀态。但是在Pentium III對SSE的實作中,浮點數寄存器又一次被新的指令集占用了,但是這一次切換運算模式不是必要的了,隻是SSE和浮點數指令不能同時進入CPU的處理線而已。
庫檔案說明
#ifndef __METHOD
#define __METHOD
void ScaleValue1(float *pArray, DWORD dwCount, float fScale);//乘法
void ScaleValue2(float *pArray, DWORD dwCount, float fScale);
void Add1(float *pArray, DWORD dwCount, float fScale);//加法
void Add2(float *pArray, DWORD dwCount, float fScale);
void Sqrt1(float *pArray, DWORD dwCount, float fScale);//平方
void Sqrt2(float *pArray, DWORD dwCount, float fScale);
void Min1(float *pArray, DWORD dwCount, float fScale);//最小值
void Min2(float *pArray, DWORD dwCount, float fScale);//最小值
void Max1(float *pArray, DWORD dwCount, float fScale);//最小值
void Max2(float *pArray, DWORD dwCount, float fScale);//最小值
void And1(float *pArray, DWORD dwCount, float fScale);//與操作
void And2(float *pArray, DWORD dwCount, float fScale);//與操作
#endif
#include <xmmintrin.h>
#include <Windows.h>
#include <math.h>
void ScaleValue1(float *pArray, DWORD dwCount, float fScale)//乘法
{
DWORD dwGroupCount = dwCount/;
__m128 e_Scale = _mm_set_ps1(fScale);//設定所有4個值為同一值
for (DWORD i=; i<dwGroupCount; i++)
{
*(__m128*)(pArray + i*) = _mm_mul_ps( *(__m128*)(pArray + i*),e_Scale);
}
}
void ScaleValue2(float *pArray, DWORD dwCount, float fScale)
{
for (DWORD i =; i<dwCount; i++)
{
pArray[i] *= fScale;
}
}
void Add1(float *pArray, DWORD dwCount, float fScale)//加法
{
DWORD dwGroupCount = dwCount/;
__m128 e_Scale = _mm_set_ps1(fScale);//設定所有4個值為同一值
for (DWORD i=; i<dwGroupCount; i++)
{
*(__m128*)(pArray + i*) = _mm_add_ps( *(__m128*)(pArray + i*),e_Scale);
}
}
void Add2(float *pArray, DWORD dwCount, float fScale)
{
for (DWORD i =; i<dwCount; i++)
{
pArray[i] += fScale;
}
}
void Sqrt1(float *pArray, DWORD dwCount, float fScale)//平方
{
DWORD dwGroupCount = dwCount/;
__m128 e_Scale = _mm_set_ps1(fScale);//設定所有4個值為同一值
for (DWORD i=; i<dwGroupCount; i++)
{
*(__m128*)(pArray + i*) = _mm_sqrt_ps(e_Scale);
}
}
void Sqrt2(float *pArray, DWORD dwCount, float fScale)
{
for (DWORD i =; i<dwCount; i++)
{
pArray[i] = sqrt(fScale);
}
}
void Min1(float *pArray, DWORD dwCount, float fScale)//最小值
{
DWORD dwGroupCount = dwCount/;
__m128 e_Scale = _mm_set_ps1(fScale);//設定所有4個值為同一值
for (DWORD i=; i<dwGroupCount; i++)
{
*(__m128*)(pArray + i*) = _mm_min_ps( *(__m128*)(pArray + i*),e_Scale);
}
}
void Min2(float *pArray, DWORD dwCount, float fScale)
{
for (DWORD i =; i<dwCount; i++)
{
pArray[i] = (pArray[i]>fScale? fScale : pArray[i]);
}
}
void Max1(float *pArray, DWORD dwCount, float fScale)//最大值
{
DWORD dwGroupCount = dwCount/;
__m128 e_Scale = _mm_set_ps1(fScale);//設定所有4個值為同一值
for (DWORD i=; i<dwGroupCount; i++)
{
*(__m128*)(pArray + i*) = _mm_max_ps( *(__m128*)(pArray + i*),e_Scale);
}
}
void Max2(float *pArray, DWORD dwCount, float fScale)
{
for (DWORD i =; i<dwCount; i++)
{
pArray[i] = (pArray[i]<fScale? fScale : pArray[i]);
}
}
void And1(float *pArray, DWORD dwCount, float fScale)//與操作
{
DWORD dwGroupCount = dwCount/;
__m128 e_Scale = _mm_set_ps1(fScale);//設定所有4個值為同一值
for (DWORD i=; i<dwGroupCount; i++)
{
*(__m128*)(pArray + i*) = _mm_and_ps( *(__m128*)(pArray + i*),e_Scale);
}
}
void And2(float *pArray, DWORD dwCount, float fScale)
{
for (DWORD i =; i<dwCount; i++)
{
pArray[i] = (int)(pArray[i]) & (int)(fScale);
}
}
采用SSE和不采用SSE的數學計算操作速度對比:
#include <xmmintrin.h>
#include <Windows.h>
#include <iostream>
#include "Method.h"
using namespace std;
#define ARRAYCOUNT 1000
#define COUNTSIZE 10000
class CTimer
{
public:
__forceinline CTimer(void)
{
QueryPerformanceFrequency(&m_Frequency);// 擷取時鐘周期
QueryPerformanceCounter(&m_StartCount);// 擷取時鐘計數
}
__forceinline void Reset(void)
{
QueryPerformanceCounter(&m_StartCount);
}
__forceinline double End(void)
{
QueryPerformanceCounter(&m_EndCount);
return ( m_EndCount.QuadPart - m_StartCount.QuadPart )*/m_Frequency.QuadPart;
}
private:
LARGE_INTEGER m_Frequency;
LARGE_INTEGER m_StartCount;
LARGE_INTEGER m_EndCount;
};
int __cdecl main()
{
float __declspec(align())Array[ARRAYCOUNT];
//__declspec(align(16))做為數組定義的修釋符,這表示該數組是以16位元組為邊界對齊的,
//因為SSE指令隻能支援這種格式的記憶體資料
memset(Array, , sizeof(float)*ARRAYCOUNT);
CTimer t;
double dTime;
//乘法
cout<<"乘法:"<<endl;
t.Reset();
for (int i=; i<COUNTSIZE; i++)
{
ScaleValue1(Array, ARRAYCOUNT, );
}
dTime = t.End();
cout<<"Use SSE: "<<dTime<<"毫秒"<<endl;
t.Reset();
for (int i=; i<COUNTSIZE; i++)
{
ScaleValue2(Array, ARRAYCOUNT, );
}
dTime = t.End();
cout<<"Not Use SSE: "<<dTime<<"毫秒"<<endl;
//加法
cout<<"加法:"<<endl;
t.Reset();
for (int i=; i<COUNTSIZE; i++)
{
Add1(Array, ARRAYCOUNT, );
}
dTime = t.End();
cout<<"Use SSE: "<<dTime<<"毫秒"<<endl;
t.Reset();
for (int i=; i<COUNTSIZE; i++)
{
Add2(Array, ARRAYCOUNT, );
}
dTime = t.End();
cout<<"Not Use SSE: "<<dTime<<"毫秒"<<endl;
//平方
cout<<"平方:"<<endl;
t.Reset();
for (int i=; i<COUNTSIZE; i++)
{
Sqrt1(Array, ARRAYCOUNT, );
}
dTime = t.End();
cout<<"Use SSE: "<<dTime<<"毫秒"<<endl;
t.Reset();
for (int i=; i<COUNTSIZE; i++)
{
Sqrt2(Array, ARRAYCOUNT, );
}
dTime = t.End();
cout<<"Not Use SSE: "<<dTime<<"毫秒"<<endl;
//最小值
cout<<"最小值:"<<endl;
t.Reset();
for (int i=; i<COUNTSIZE; i++)
{
Min1(Array, ARRAYCOUNT, );
}
dTime = t.End();
cout<<"Use SSE: "<<dTime<<"毫秒"<<endl;
t.Reset();
for (int i=; i<COUNTSIZE; i++)
{
Min2(Array, ARRAYCOUNT, );
}
dTime = t.End();
cout<<"Not Use SSE: "<<dTime<<"毫秒"<<endl;
//最大值
cout<<"最大值:"<<endl;
t.Reset();
for (int i=; i<COUNTSIZE; i++)
{
Max1(Array, ARRAYCOUNT, );
}
dTime = t.End();
cout<<"Use SSE: "<<dTime<<"毫秒"<<endl;
t.Reset();
for (int i=; i<COUNTSIZE; i++)
{
Max2(Array, ARRAYCOUNT, );
}
dTime = t.End();
cout<<"Not Use SSE: "<<dTime<<"毫秒"<<endl;
//與操作
cout<<"與操作:"<<endl;
t.Reset();
for (int i=; i<COUNTSIZE; i++)
{
And1(Array, ARRAYCOUNT, );
}
dTime = t.End();
cout<<"Use SSE: "<<dTime<<"毫秒"<<endl;
t.Reset();
for (int i=; i<COUNTSIZE; i++)
{
And2(Array, ARRAYCOUNT, );
}
dTime = t.End();
cout<<"Not Use SSE: "<<dTime<<"毫秒"<<endl;
system("pause");
return ;
}