Right now the linear algebra library we use -ublas- has the same behaviour as Eigen for BLAS1 type expressions. So it tries to generate optimal (non-SSE) code. Only for BLAS2 and 3 we fall back to the ATLAS-routines which has the same performance as Eigen on the interesting problem sizes.
//small edit In the end it is not so interesting whether the BLAS1-type expressions are fast as they make up < 1% of run time performance. The big chunks are the data processing inside the matrix-matrix multiplications of the Neural Networks and similar entities.
Another thing about code generation, I am also using a hacked version of Eigen as well in a project I'm working on that can do the tanh and derivative of the tanh so the NN activations go quite abit faster since you can generate vectorized code for the whole calculation that will visit the memory location exactly once. While true the calculation of the weight updates is the most time spent, I saw 3-4x speedup in the activation code doing it in a single operation due to better memory access patterns and less loop iterations. Better memory access patterns can also have synergistic effects on other code because there is less cache pollution happening. By being fast and loose and introducing a few other copies of the matrix data in my case, my performance falls off a cliff when it no longer fits in the cpu cache nicely. 10x difference in the particular case I am remembering.
As always performance is part art, part science and perhaps it won't matter as much for the general case, but for my specific implementation and my matrix sizes Eigen has made a measurable difference for me compared to other solutions.
Right now I would say that the main focus of shark is research oriented. That is we want to be fast but also modular so that we can still easily exchange different aspects of the algorithms with our own work. As these goals sometime clash, it is hard to claim that we are the fastest, simply because there is for nearly every algorithm some way to improve when you know exactly which combination of model, loss function and training algorithm you use. But we are (hopefully) reasonably fast and certainly want to improve.
First of all, we are glad that our library is discussed on this board! We are happy for every feedback we can get!
Regarding Performance: we try to get the key algorithms as fast as possible. And for the hardest parts we rely not on ublas, but use optional bindings to ATLAS. Speed was one of the key design criteria. We hope that we achieved that. Clearly this is no guarantee that every algorithm is fast, but in this case: just add a ticket!
Please bear in mind, that Shark is still in beta stage, and we are heavily developing it (I am right now working on the family of multi class SVMs). So for example parallelism using OpenMP is not fully integrated.