DGETRF Performance issue on Xeon Phi

Open discussion for MAGMA

DGETRF Performance issue on Xeon Phi

Postby mathieudubois » Tue Apr 16, 2013 8:40 am

I have been testing the performance of the MAGMA library for Xeon Phi, and I am a little bit disappointed with the performance. In particular, I have benchmarked the DGETRF function and compared to the performance using Intel MKL library on standard CPU. Basically I was trying to reproduce the performance of slide 15 of presentation http://icl.cs.utk.edu/projectsfiles/mag ... MIC_03.pdf. My system is very similar to the one used for those tests, but I can only reach 200-250 GFlops on Xeon Phi, whereas on 16 Sandy Bridge cores I am close to 300 GFlops.

I would like to know if anyone can help get reproduce those results ? Thank you very much in advance for your help.

Other question, is there an implementation of DGETRI planned to be released for Xeon Phi ?

Cheers,

Mathieu
mathieudubois
 
Posts: 1
Joined: Tue Apr 16, 2013 8:24 am

Re: DGETRF Performance issue on Xeon Phi

Postby mgates3 » Thu Apr 18, 2013 11:34 am

This is a beta release. Many performance improvements are coming.

MAGMA for CUDA has getri, so it may eventually make it into MAGMA for Xeon Phi. However, it is generally both faster and more accurate to solve a linear system Ax=b using gesv (that is, getrf + getrs) than to invert the matrix and multiply (getrf + getri + gemm). There are a limited number of applications that need an explicit matrix inverse.

-mark
mgates3
 
Posts: 409
Joined: Fri Jan 06, 2012 2:13 pm


Return to User discussion

Who is online

Users browsing this forum: No registered users and 1 guest