Page 1 of 1

CUDA_DGESV crashes for more than 15k size

PostPosted: Fri Jan 20, 2012 2:46 pm
by bhamarsd
I have been using the CULA_DGESV routine for quite a while now. It runs very fast and accurately for 'A' size less than 15,200 (to be exact). Beyond this limit, it either gives the 'Data Error' or returns the erroneous results. Is there a bug or it's just the limit this CULA version can handle ?

I'm using CULA R11.

Thanks

Re: CUDA_DGESV crashes for more than 15k size

PostPosted: Mon Jan 23, 2012 6:37 am
by john
When you say "crash" do you mean a literal crash - ie your program seg. faults? I ask because you've noted erroneous data and/or error codes, which indicates to me that the routine has run without crashing.

15k x 15k happens to be right about at 1 GB. Are you certain that you're able to allocate the full matrix on the GPU? You didn't state your system configuration, so I can only hypothesize. CULA's host interface should return an out of memory status if it can't allocate the whole matrix, but the device interface obviously requires that you have already pre-allocated your matrix data.

DataError indicates a singular matrix has been encountered. It could be that your whole matrix has not been loaded to the card properly or that your data is possibly not stable to single precision at these larger sizes?

Lastly, I would strongly advise moving up from R11 to the latest version and testing on the latest to ensure this is still a problem for you.