[linalg] cling xtensor dot product results inconsistent with numpy - `nan`'s about xtensor-blas

[wolfv]

This seems to be a really weird issue. Luckily, so far I could only reproduce it in the jupyter notebook.

Using this code:

template <class C>
auto countnan(const C& c)
{
    int i = 0;
    for (auto& el : c)
    {
        if (std::isnan(el))
        {
             i += 1;
        }
    }
    return i;
}

I was able to observe a very random nondeterministic number of nans in the result.

I wonder if this is somehow related to cling.

[wolfv]

Okay, I took a second look into it.

We do import the blas libraries for cling, but we do not define HAVE_CBLAS 1 (e.g.

#define HAVE_CBLAS 1

If you define this at the top of your notebook, you should not see any nan's anymore (at least that was the case for me).

Now that hints at a problem in the generic implementation of the dot product in FLENS.

I am looking at the output from running with CXXBLAS_DEBUG 1 and I see that CXXBLAS: dotu_generic is getting called. I'll take a look if I can find something obvious.

[wolfv]

Ok, i found the root cause of the issue in this line of FLENS

for (IndexType i=0, iY=0; i<m; ++i, iY+=incY) {
    VY y_;
    dotu_generic(n, A+i*ldA, IndexType(1), x, incX, y_);
    y[iY] += alpha * y_;
}

The problem is that it's += assigning to potentially unitialized memory ... and that produces random nan's!

Let's check what the BLAS manual says.