topic 64-bit Workspace with sgesvd? in Intel® oneAPI Math Kernel Library

64-bit Workspace with sgesvd?

schmmd — Thu, 16 Apr 2009 19:46:34 GMT

Is it possible to have a 64-bit work array with sgesvd? Since the lwork parameter is a signed 32-bit integer, this does not seem possible, but maybe there is a similar function I am overlooking.

Re: 64-bit Workspace with sgesvd?

TimP — Thu, 16 Apr 2009 20:21:12 GMT

Quoting - schmmd

Is it possible to have a 64-bit work array with sgesvd? Since the lwork parameter is a signed 32-bit integer, this does not seem possible, but maybe there is a similar function I am overlooking.

The ilp64 library version (not available for ia32) presumably has a 64-bit work array. Maybe that's not what you are asking.

Re: 64-bit Workspace with sgesvd?

schmmd — Fri, 17 Apr 2009 00:30:09 GMT

Quoting - tim18

The ilp64 library version (not available for ia32) presumably has a 64-bit work array. Maybe that's not what you are asking.

I don't think this is a problem with the actual library but rather with the header file. I need to specify the length of the work array and I can only send in an "int" with sgesvd. Moreover, all the libraries use the same header files. So I think that if I use sgesvd I am forced to only send a 32-bit int.

Re: 64-bit Workspace with sgesvd?

TimP — Fri, 17 Apr 2009 03:49:19 GMT

We got a confirmation in an earlier thread of some errors in the headers for ilp64, where all integers are 64-bit.

Re: 64-bit Workspace with sgesvd?

Michael_C_Intel4 — Fri, 17 Apr 2009 11:25:30 GMT

Quoting - schmmd

Is it possible to have a 64-bit work array with sgesvd? Since the lwork parameter is a signed 32-bit integer, this does not seem possible, but maybe there is a similar function I am overlooking.

Hello,

do you mean your problem size is so huge that the space needed for work array doesn't fit to 32-bit space? Then there are several options:
a) if you're workingwith 32-bit MKL the only way is to reduce workspace amount: the minimal workspace needed is max(3*min(m, n)+max(m, n), 5*min(m,n)) of floats, I don't believe the linear size of your problem is so huge not tofit the minimal workspace into 32-bit.
b) if you're working with 64-bit MKL you may reduce your workspace according to a) in case you need only LP64 model; or you may switch to ILP64 model where lwork is 64-bit.

Or do you mean something different?

Michael.