Page 1 of 1

R14 pcula_Sgemm...

PostPosted: Thu Jun 07, 2012 12:46 pm
by modelsciences
This is really about the error I am getting:

on culaStatus == culaArgumentError
culaGetErrorInfo() returns -2

This AFIAK should tell me which argument is wrong: -2 seems a bit of an odd one...
unless we are counting backwards from the end :-)

Any ideas?

-- Simon

Re: R14 pcula_Sgemm...

PostPosted: Thu Jun 07, 2012 3:51 pm
by kyle
This is the LAPACK/BLAS standard that is mostly a relic from the Fortran roots. Negative error codes are argument errors, positive error codes are data errors.

Re: R14 pcula_Sgemm...

PostPosted: Thu Jun 07, 2012 4:01 pm
by modelsciences
LAPACK as an API seems like a relic for sure. Thanks for the tip - so I'll take a close look at arg 2 :-) now is that 1 or 0 based and do we count the pcula context? I guess not.

-- Simon

Re: R14 pcula_Sgemm...

PostPosted: Fri Jun 08, 2012 4:50 am
by modelsciences
Well the answer was that this LAPACK seems case sensitive as to the N|T flags - n|t will not do!
This is a new on me - I have always used lowercase flags before!

BTW the parameter numbering starts from 1 (or -1) and DOES include the pculaConfig* argument.

-- Simon

Re: R14 pcula_Sgemm...

PostPosted: Fri Jun 08, 2012 4:59 am
by kyle
Remember R14's pCULA is very much in an early ALPHA state. Bugs should be expected!

That being said, the current internal version (R15) has vastly improved stability, speed, and scalability.

Re: R14 pcula_Sgemm...

PostPosted: Fri Jun 08, 2012 6:22 am
by modelsciences
We'll be ready for R15 :-) I am just prototyping some new codes that will really leverage this approach to multi-gpu.

Keep up the good work!

-- Simon

Re: R14 pcula_Sgemm...

PostPosted: Fri Jun 08, 2012 1:23 pm
by kyle
Great! We'd be happy to hear any interfacing/usage feedback/suggestions while we are still working out the details.