Re: [eigen] Blas performance on mapped matrices

[Jitse Niesen <jitse@xxxxxxxxxxxxxxxxx>]

On Mon, 9 Jan 2012, Sameer Agarwal wrote:

> We are in the process of a significant code migration from eigen2 to
> eigen3. The code uses Eigen::Map to map chunks of memory into RowMajor
> matrices and operates on them. The primary operation is of the form
>
> A.block(r, c, size1, size2) -= B * C;
> [...]
> Moving from eigen2 to eigen3 has resulting in a 30% performance
> regression. Has something changed significantly in the way Eigen3
> handles mapped matrices, or about the structure of matrix-matrix
> multiplication in Eigen3 that would cause this?

Gael did a lot of work on matrix-matrix multiplication in between Eigen2 
and Eigen3, so I guess it may be related to that.

You say in a later email that you know your instruction can be rewritten 
as

   A.block<size1, size2>(r, c) -= B * C;

if the size of the block is known at compile time, which seems to be the 
case in the example you're concentration on (becauase the size of B and C 
are known). I assume you tested that this is not the cause of the 
performance regression.

You should also try writing it as

   A.block<size1, size2>(r, c).noalias() -= B * C;

(assuming that is valid).

But I guess it's something more complicated. Another knob you can play 
with is the EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD constant, which is used 
to decide which implementation to use for matrix multiplication. Try 
changing it to 16 (anything larger than 9 will do for your example).


Jitse