Performance Comparison to Intel MKL (MATLAB) about blasfeo HOT 3 OPEN

@giaf , any thought on this?
Could you run your benchmarks on Linux with Multi Threaded version of MKL?

from blasfeo.

Hi @RoyiAvital
it's difficult for me to judge without playing around myself, there are a lot of factors affecting performance, and also MEX introduces some overhead, it's not for free.
Could you also upload the source code for the MEX wrapper?

On top of that, MKL is not a sitting dog, they have a large team and, according to the MKL release notes, they have been improving their small scale performance too.
While here man power is very limited and divided among multiple projects :p
I think benchmarks on blasfeo.syscop.de have not been updates since the BLASFEO BLAS API paper was initially submitted, so I'll try to find some time to repeat them.

About comparison with multi-threaded MKL, it's not an apple-to-apple comparison, as MKL performance would scale depending on the number of available cores (and then the CPU model), while BLASFEO's one would not.
I eventually want to add multi-thread capabilities to BLASFEO too, at that point such comparison would make sense in my opinion.

But as of now, BLASFEO is designed with the aim of providing fast single-threaded routines for matrices fitting in cache.
There are multiple applications where this is needed, as e.g. in the PLASMA project.
And especially (for applications in our group) to provide the linear algebra framework for the implementation of embedded optimization algorithms for optimal control software such as HPIPM and acados.
This is the main aim in development.

from blasfeo.

@giaf ,
I know the idea is to have Single Threaded performance.
Yet I think comparing to Multi Threaded MKL will give a break point where one should use BLASFEO and MKL.
My thought it worth you add those wasn't to show BLASFEO in negative light, on the contrary, it will probably show how close you get even with a single thread.

By the way, MEX indeed has an overhead. But remember MATLAB also calls MKL using its general purpose DLL which has the same overhead to the least (Probably more). Also MATLAB uses the CNR channel of MKL which has lower performance (Yet reproducible results).

This is the MEX file (Change to c postfix):

BLASFEODGEMM.txt

Pay attention it is not fully working and there is no validation of the input in order to compare pure performance (It works for square matrices).

from blasfeo.