question about new QR functions geqp3 and geqrt

[Azua_Garcia__Giovann]

Hello,

First off, how can I tell the exact MKL version installed? I know I have the latest (it has LAPACK 3.4.0) but it becomes confusing to tell after applying a few updates. I see under /opt/intel the following which looks fishy with the "2011":

composerxe -> composer_xe_2011_sp1/
drwxr-xr-x 3 root root 4096 2012-05-05 17:44 composer_xe_2011_sp1/
drwxr-xr-x 8 root root 4096 2012-05-05 17:44 composer_xe_2011_sp1.10.319/
drwxr-xr-x 14 root root 4096 2012-05-04 01:04 composer_xe_2011_sp1.9.293/

Anyway I'm having trouble making these newly included QR functions work and I suspect it could be bugs? The problems I am having are the following:

dgeqrt: No matter how I call this function, I always get the following error
MKL ERROR: Parameter 3 was incorrect on entry to DGEQRT.
LAPACKE_dgeqrt failed with info = -4

dgeqp3: I get the R not solving a test case problem that gets solved using the classic dgeqrf. I will prepare a demonstration test case in isolation for this.

Many TIA,
Best regards,
Giovanni

[Gennady_F_Intel]

- regarding exact version installed: please see the file/opt/intel/composer_<..>/Documentation/mklsupport

you can see there smth like in my case --Package ID: l_mkl_10.3.10.319 l_ccompxe_2011.10.319 l_fcompxe_2011.10.319

also - you can usethemkl_get_version(MKLVersion*pVersion) function to obtain anMKLVersionstructure that contains the version information.

- regarding dgeqrt's error - pls give us the example?

[Azua_Garcia__Giovann]

Hi Gennady,

Thank you! I figured the dgeqrt's error, the problem was it required the NB >= min(M, N), I was fixing it to 64 and didn't work. I managed to get hold of the LAPACK online doc and this pre-condition is stated there. The mklman pdf doesn't mention this though.

I will prepare a small example for the geqp3 case. It basically fails to factorize/solve a simple overdetermined system.

Thanks again!
Best regards,
Giovanni

[mecej4 · Accepted solution]

The MKL manual that got installed with MKL 10.3.10 does state the requirement that "min(m, n nb 1" -- it is missing a right parenthesis after ", n", but the essential information is there to see. It is possible that you are looking at an older manual in which the requirement was not stated. That is a problem with having so many manuals lying around.

MKL does not come with an example for exercising geqp3, but there is such an example provided by NAG with their FL libraries that is fairly easy to adapt to use with MKL instead. That example involves a 6 X 5 matrix and MKL gives results that are very close to the NAG FL results given there.