syevd performance

Open discussion for MAGMA library (Matrix Algebra on GPU and Multicore Architectures)
Post Reply
Posts: 5
Joined: Tue Jan 10, 2012 6:34 am

syevd performance

Post by yariveis » Tue Jan 17, 2012 7:25 am


I am running the testing_dsyevd.cpp from magma_1.0.0-rc5 version.
I am running on Windows 7 64 bit with Geforce 540M, 2Giga memory.

When comparing the results from the CPU runs to the GPU runs I see that there is an improvment of factor 2 between the runs (on matrices with size 2000 and higher).
For instance I see that for matrix size 3072 I got the result on the CPU within 123 seconds while on the GPU it took 55 seconds.
The same task on matlab took 44 seconds.

Does anyone knows what method does matlab uses to compute the eig problem?
Is there a way to improve the performace? Maybe some tuning in the code?
Would changing to magma 1.1.0 can make any improvment here?


Posts: 18
Joined: Tue Jan 25, 2011 8:20 pm

Re: syevd performance

Post by brom » Wed Jan 18, 2012 10:18 am

MATLAB uses Intel's MKL which has a special algorithm called successive band reduction for the symmetric eigenvalue factorization step.

Post Reply