Weighted sum in implemented in Xtensor runs about x4 slower than native C++

[davidsajman]

Hi,

Xtensor implementation of weighted sum:
double xaverage(const pyvector& x, const pyvector& weights) {

struct timeval start, end;
gettimeofday(&start, NULL);

double res = xt::average(x, weights)[0];

gettimeofday(&end, NULL);
float rtime = (end.tv_sec-start.tv_sec)*1000.0f+(end.tv_usec-start.tv_usec)/1000.0f;
printf("weighted average run time: %f ms\n",  rtime);

return res;

}

C++ implementation of weighted sum with Pybind:
double wavg(py::array_t& vals, py::array_t& weights)
{

struct timeval start, end;
gettimeofday(&start, NULL);

auto info = vals.request();
double* v = (double*)info.ptr;
double* w = (double*)weights.request().ptr;
double sum=0;
double wsum=0;
for (int i=0; i<info.shape[0]; i++)
{
    sum+=v[i]*w[i];
    wsum+=w[i];
}
double res = sum/wsum;

gettimeofday(&end, NULL);
float rtime = (end.tv_sec-start.tv_sec)*1000.0f+(end.tv_usec-start.tv_usec)/1000.0f;
printf("wavg run time: %f ms\n",  rtime);

return res;

}

Does somebody have an idea why Xtensor is so significantly slower?

[emmenlau]

Is it possible that you did not activate the SIMD instructions in xtensor? See https://xtensor.readthedocs.io/en/latest/build-options.html#external-dependencies

[tdegeus]

Also it could help to ensure that x and weights are xt::xtensor<T, N> if you can

[davidsajman]

Regarding XSIMD - it is set by the conan configuration in my environment. I'll recheck it is on.

I checked that inputs are xt::xtensor<T, N> - same result.

[JohanMabille]

We haven't plugged SIMD in reducers yet. My assumption is that your code is automagically vectorized by the compiler while the reducer one is not.