Bugzilla – Bug 1166
MatrixXd block, VectorXd and noalias
Last modified: 2016-02-19 01:41:43 UTC
Created attachment 653 [details]
While investigating an error in our code relying on Eigen, I noticed the following behavior. Given a MatrixXd a, the result of this operation on a block b of a MatrixXd m is wrong:
b.noalias() = 2. * x.transpose() * a;
when noalias() is used and x is a VectorXd. As a result, the block b of the matrix contains the wrong data (misplaced in memory). However, if x is a MatrixXd with the proper size, things work normally. I attached a repro code that shows the issue. I tested in on Eigen 3.2.7 and 3.3-beta1.
Am I doing something wrong here?
I confirm. The problem is when extracting the inner-stride of the destination using:
which is not necessarily correct when the destination expression is not a vector expression at compile-time, as with m.block(1, 0, 1, 2). This works fine with m.row(1), which is the workaround you should follow.
On Eigen's side, a quick fix would be enforce the destination to be a vector by calling:
This should not produce any runtime overhead.
Duly noted, thanks for the quick answer!
Summary: Bug 1166: fix shortcomming in gemv when the destination is not a vector at compile-time.
btw, thanks for find that one.
Well thanks for the quick fix (and maintenance release), it's greatly appreciated!