求点集平均值的指令集加速例子

最近学指令集加速，比较好奇求平均值能不能用指令集加速。于是做了这么一个测试。先说一下结论：不要看普通C++代码循环内部调用了operator+=()函数就觉得慢，其实在Release版优化的情况下效率和使用指令集加速差异在2%左右。所以像求平均值这种简单的运算不需要手工调用指令集优化，编译器可以做的很好。下述代码基于VS2017、Qt5.9和OpenCV430，CPU型号是Intel Core i5-7400。下面是代码：

static double sset = 0;
static double nonet = 0;
static double opencvt = 0;

void main()
{
    vector<Point2f> points1(10000);
    vector<Point2f> points2(10000);
    vector<Point2f> points3(10000);
    randu(points1, 0, 10);
    randu(points2, 0, 10);
    randu(points3, 0, 10);

    int64 t1, t2;
    Point2f avr;

    for (int x = 0; x < 100000; x++)
    {
        t1 = getTickCount();
        for (auto& item : points1)
        {
            avr += item;
        }
        avr /= 10000.0f;
        t2 = getTickCount();
        nonet += (t2 - t1) / getTickFrequency() * 1000;

        t1 = getTickCount();
        __m128 su128 = _mm_setzero_ps();
        for (int i = 0; i < 10000; i += 2)
        {
            __m128 mmx = _mm_loadu_ps((float*)&points2[i]);
            su128 = _mm_add_ps(su128, mmx);
        }
        float e1, e2;
        _MM_EXTRACT_FLOAT(e1, su128, 0);
        _MM_EXTRACT_FLOAT(e2, su128, 2);
        avr.x = (e1 + e2) / 10000;
        _MM_EXTRACT_FLOAT(e1, su128, 1);
        _MM_EXTRACT_FLOAT(e2, su128, 3);
        avr.y = (e1 + e2) / 10000;
        t2 = getTickCount();
        sset += (t2 - t1) / getTickFrequency() * 1000;

        t1 = getTickCount();
        __m256 su256 = _mm256_setzero_ps();
        for (int i = 0; i < 10000; i += 4)
        {
            __m256 mmx = _mm256_loadu_ps((float*)&points3[i]);
            su256 = _mm256_add_ps(su256, mmx);
        }
        __m128 mml = _mm256_extractf128_ps(su256, 0);
        __m128 mmh = _mm256_extractf128_ps(su256, 1);
        __m128 res = _mm_add_ps(mml, mmh);
        _MM_EXTRACT_FLOAT(e1, res, 0);
        _MM_EXTRACT_FLOAT(e2, res, 2);
        avr.x = (e1 + e2) / 10000;
        _MM_EXTRACT_FLOAT(e1, res, 1);
        _MM_EXTRACT_FLOAT(e2, res, 3);
        avr.y = (e1 + e2) / 10000;
        t2 = getTickCount();
        opencvt += (t2 - t1) / getTickFrequency() * 1000;
    }
    qDebug() << "NONE(ms):" << nonet;
    qDebug() << "SSE(ms):" << sset;
    qDebug() << "AVX(ms):" << opencvt;
}

下面是运行几次的输出。可以看出指令集加速毫无作用甚至适得其反：

NONE(ms): 1.6826
SSE(ms): 1.6635
AVX(ms): 1.6978
NONE(ms): 1.7824
SSE(ms): 2.0012
AVX(ms): 1.8098
NONE(ms): 1.6637
SSE(ms): 1.817
AVX(ms): 1.7181
NONE(ms): 1.7506
SSE(ms): 2.2308
AVX(ms): 1.7104
NONE(ms): 1.7499
SSE(ms): 2.8038
AVX(ms): 1.783