opencv转置第一次需要很长时间

Question

我使用opencv 3.1，并通过visual studio 2015更新3在windows 10中启用intel TBB。第一个转置时间需要100ms，而另一个转置只需要0.02-0.05ms。任何人都知道为什么第一次转置对于1*1矩阵需要这么多时间。

double ts = time_measure("start", 0);
Mat_<uchar> A = (Mat_<uchar>(1, 1) << 1);
Mat at = A.t();
cout << "transpose Times needed : " << time_measure("end", ts) * 1000 << " ms " << endl;
for (int i = 0; i < 10; i++) {
    ts = time_measure("start", 0);
    Mat_<uchar> B = (Mat_<uchar>(1, 1) << 1);
    Mat bt = B.t();
    cout << "transpose Times needed : " << time_measure("end", ts) * 1000 << " ms " << endl;
}

double time_measure(const string mode, double ts) {
    double t = 0.0;
    if (mode == "start") {
       t = (double)getTickCount();
    }
    else {
       t = ((double)getTickCount() - ts) / getTickFrequency();
    }
    return t;
}
The output
transpose A Times needed : 112.062 mstranspose B Times needed : 0.0337221 ms
transpose B Times needed : 0.0205265 ms
transpose B Times needed : 0.0195491 ms
transpose B Times needed : 0.0283461 ms
transpose B Times needed : 0.0234589 ms
transpose B Times needed : 0.0298123 ms
transpose B Times needed : 0.0249251 ms
transpose B Times needed : 0.0283461 ms
transpose B Times needed : 0.0273687 ms
transpose B Times needed : 0.02688 ms

Answer 1

我没有启用TBB，但衡量性能的方式似乎有问题：

1. 不包括创建矩阵的时间
2. 使用足够大的矩阵。无论如何，转置1x1矩阵是没有意义的
3. 不要将字符串用于布尔值

你可以试试这样的方法，然后请告诉我你的执行时间：

double time_measure(bool start, double ts) {
    double t = 0.0;
    if (start) {
        t = (double)getTickCount();
    }
    else {
        t = ((double)getTickCount() - ts) / getTickFrequency();
    }
    return t;
}
int main()
{
    for (int i = 0; i < 10; i++) {
        // 1000 x 1000 random matrix
        Mat_<uchar> B(1000, 1000);
        randu(B, 0, 256);
        double ts = time_measure(true, 0);
        Mat bt = B.t();
        cout << "transpose Times needed : " << time_measure(false, ts) * 1000 << " ms " << endl;
    }
    getchar();
    return 0;
}

Answer 2

感谢您的评论。我试过你的代码，矩阵的创建时间是固定的，第一个转置时间也比另一个转置花费更多的时间。以下是测试结果。我修改代码以打印矩阵创建时间和转换时间。

int _tmain(int argc, _TCHAR* argv[]) {
    for (int i = 0; i < 10; i++) {
        double ts = time_measure(true, 0);
        // 1000 x 1000 random matrix
        Mat_<uchar> B(1000, 1000);
        randu(B, 0, 256);
        cout << "create matrix Times needed : " << time_measure(false, ts) * 1000 << " ms " << endl;
        ts = time_measure(true, 0);
        Mat bt = B.t();
        cout << "transpose Times needed : " << time_measure(false, ts) * 1000 << " ms " << endl;
    }
}
double time_measure(bool start, double ts) {
double t = 0.0;
if (start) {
    t = (double)getTickCount();
}
else {
    t = ((double)getTickCount() - ts) / getTickFrequency();
}
return t;
}

输出如下所示：

create matrix Times needed : 49.3267 ms
transpose Times needed : 427.299 ms
create matrix Times needed : 51.8431 ms
transpose Times needed : 0.889971 ms
create matrix Times needed : 51.8084 ms
transpose Times needed : 0.718917 ms
create matrix Times needed : 52.4946 ms
transpose Times needed : 0.742376 ms
create matrix Times needed : 45.5454 ms
transpose Times needed : 0.705721 ms
create matrix Times needed : 45.218 ms
transpose Times needed : 0.70621 ms
create matrix Times needed : 44.5748 ms
transpose Times needed : 0.713541 ms
create matrix Times needed : 46.2501 ms
transpose Times needed : 0.68715 ms
create matrix Times needed : 45.153 ms
transpose Times needed : 0.663691 ms
create matrix Times needed : 44.1892 ms
transpose Times needed : 0.584028 ms