stream.cpp

获得多核硬件所有的相关信息。汇编源代码。Visual Studio2003、2005环境皆可。速度快。

//---------------------------------------------------------------------------
//
// Stream.cpp: STREAM Routines Implementation
//
//---------------------------------------------------------------------------

#include "StdAfx.h"
#include "stream.h"

void __cdecl StreamCopy(void* dst, void* src, DWORD count, DWORD repeat)
{
    DWORD cnt = count >> 3;
    for (DWORD i = 0; i < repeat; ++i)
    {
        for (DWORD j = 0; j < cnt; ++j)
        {
            ((double *)dst)[j] = ((double *)src)[j];
        }
    }
}

void __cdecl StreamScale(void* dst, void* src, DWORD count, DWORD repeat)
{
    DWORD cnt = count >> 3;
    double f = 3.0;
    for (DWORD i = 0; i < repeat; ++i)
    {
        for (DWORD j = 0; j < cnt; ++j)
        {
            ((double *)dst)[j] = ((double *)src)[j] * f;
        }
    }
}

void __cdecl StreamAdd(void* dst, void* src1, void* src2, DWORD count, DWORD repeat)
{
    DWORD cnt = count >> 3;
    for (DWORD i = 0; i < repeat; ++i)
    {
        for (DWORD j = 0; j < cnt; ++j)
        {
            ((double *)dst)[j] = ((double *)src1)[j] + ((double *)src2)[j];
        }
    }
}

void __cdecl StreamTriad(void* dst, void* src1, void* src2, DWORD count, DWORD repeat)
{
    DWORD cnt = count >> 3;
    double f = 3.0;
    for (DWORD i = 0; i < repeat; ++i)
    {
        for (DWORD j = 0; j < cnt; ++j)
        {
            ((double *)dst)[j] = ((double *)src1)[j] + ((double *)src2)[j] * f;
        }
    }
}