zhpgvx and zhegvx

[toxa81]

Dear MKL developers!
Seems like this two important subroutines zhpgvx and zhegvx are not multi-threaded. Is this true? I run a couple
of tests for different OMP_NUM_THREADS and got no performance impact. (The same test for zgemm works fine and I get nice scaling with increasing of the number of threads).

Cheers,
Anton.


Update: zhegvx scales, but not perfectly (that was my error during the tests). I get double speed on 4 cores comparing with serial version. But zhpgvx doesn't scale.

[Michael_C_Intel4]

Anton,

it's rather difficult to get a perfect speed-up for zhegvx, so double speed on 4 cores is OK for this routine. zhpgvx is poor scaled indeed (almost no scaling), it can be improved.

Actually, we recommend to use unpacked storage routines rather than packed ones (like zhegvx instead of zhpgvx) for better performance if you have enough memory to allocate the whole matrix.

Thank you for paying attention to the MKL scaling issues.

Michael.