Hi,

I'm new to Eigen, I've used it for the past 2 weeks now. I am aware that Eigen provides parallelization for a number of algorithms as stated here. However sparse matrix-vector products, for example, are not among them. Is there a mechanism which allows to introduce parallelism to Eigen? How difficult would it be to implement a general parallelization mechanism which works on more complex expressions using lazy evaluation? (e.g. expressions like

Code: Select all

 mat = mat * vec + vec.transpose() * vec 

where vec and mat are sparse.). Alternatively, how would Eigen's current parallelization mechanism handle the example above, if now vec and mat are dense and openMP enabled so that the matrix-metrix(matrix-vector) is parallelized?

Cheers,
Teo

For sparse*vec, this is done in the devel branch (or 3.3-beta1), see respective doc: https://eigen.tuxfamily.org/dox-devel/T ... ading.html

Your exemple will fail, you are adding a vector to a matrix. Anyway, matrix products are not lazy operations. For efficiency reasons, they are computed within temporaries or directly within the destination matrix if possible. For instance, in the dense world, and using eigen 3.3, the following:

m5.noalias() = 3*m1*m2 - m3*m4;

will be evaluated as:

m5.setZero();
gemm(3,m1, m2, m5);
gemm(-1,m3, m4, m5);

where gemm stands for an internal routine similar to BLAS's gemm ones.

Dense matrix-vector products are not parallelized yet.