Issue with xt:stddev being very slow

[kritchie]

Hi, first of all thanks for the hard work on the library. It's pretty cool.

I'm using the xtensor-python but I wasn't sure to post it here of in xtensor-python.
I figured it would be more relevant here.

I have an issue with the stddev function. To keep it simple I have this code which takes as input a Numpy array of shape [5000, 1] (so basically a vector of 5K elements) and computes the mean and the standard deviation of that vector of doubles.

xt::pyarray<double> process(xt::pyarray<double>& m)
{

    xt::pyarray<double> features = {0., 0.};

    features[0] = xt::mean(m)(0, 0);
    features[1] = xt::stddev(m)(0, 0);

    return features;
}

My issue is that this code runs pretty fast when I comment the xt:stddev line (computing only the xt::mean).

features[0] = xt::mean(m)(0, 0);
// features[1] = xt::stddev(m)(0, 0);

But it's extremely slow (by a factor of 100x) when I left it uncommented. So I don't understand why adding the computation of the stddev is so slow.

There's a high chance that I'm doing something wrong here but I'm not sure what exactly. Is there a different and specific way in which the stddev method is supposed to be called ?

[JohanMabille]

Your code is fine, that's a performance issue that we fixed today ;)
The issue was because we didn't precompute the mean in the variance implementation (see https://github.com/xtensor-stack/xtensor/pull/1627/files?file-filters%5B%5D=.cpp&file-filters%5B%5D=.hpp#diff-eb049062fda2bc01f13cfe4938bc8a7eL2090)

[kritchie]

Thanks for the quick reply !

I've pulled the updates from the master branch and I've recompiled the xtensor library.

Unfortunately, I still get the same result with the stddev being slower (by a large margin) compared to the mean.

[JohanMabille]

What is the performance hit in percentage?

[chrizbee]

@JohanMabille
I have encountered the same issue. I measured some timings for my case.

The xarray to perform the stddev on is a flattened double array with 18426 elements.

xt::xarray<double> arr = xt::flatten(lums);

XTensors stddev algorithm takes ~530ms

double stddev = xt::stddev(arr)();

My own stddev implementation with STL functions takes ~46us

std::vector<double> diff(arr.size());
std::transform(arr.begin(), arr.end(), diff.begin(), [mean](double x) { return x - mean; });
double sum = std::inner_product(diff.begin(), diff.end(), diff.begin(), 0.0);
double stddev = sqrt(sum / arr.size());

So stddev with stl functions is currently just 11522 times faster than xtensor's :)