Page 1 of 1

testing_dsyr2k_mgpu and use of nb

Posted: Mon Sep 01, 2014 6:11 am
by roalmar2

I am taking times of ejecution and Gflops for this function (dsyr2k).

Which is the main purpose of the function and how works the option nb (size of block)? Is the way of computacion, or packet sending ...?

All the information is wellcoming ;)

Thanks for all.

Re: testing_dsyr2k_mgpu and use of nb

Posted: Fri Sep 19, 2014 3:49 pm
by haidar
Sorry for the delay I was away.
magmablas_dsyr2k_mgpu has been used internally within the Eigenvalues routines.
it has some assumption (for that it is not LAPACK compliant, for example if you use the routine with "L" lower case, however it touches all the nb x nb diagonal blocks)
nb is the width of the gemm used internally (dsyr2k is represented as a bench of dgemm to get higher performance) usually nb=512 to 1024 is enough to get performance.
Please let me know if you requires any further information.