topic Re: Numerical differences in DGELSS between identical calls in Intel® oneAPI Math Kernel Library

Numerical differences in DGELSS between identical calls

davidjward30 — Wed, 18 Aug 2004 18:52:05 GMT

When calling DGELSS multiple times with the same arguments, I find that the results of the FIRST call differ ever so slightly from subsequent calls which are all indentical. Typically, the differences are in the last couple of digits.

I have not found such problems with any other MKL functions.

I am using MKL7.0 on Windows XP Pro, Pentium IV 3Ghz. Visual C++ 7.1

Perhaps MKL is altering the floating point precision via the _controlfp / control87 functions?

Thanks.

David.

Re: Numerical differences in DGELSS between identical calls

TimP — Wed, 18 Aug 2004 20:22:38 GMT

Assuming that you are using the P4 library, are you passing data with identical alignments? If not, supposing that you are declaring your data in C, try __declspec(align 16) declarations.

Re: Numerical differences in DGELSS between identical calls

davidjward30 — Wed, 18 Aug 2004 23:15:03 GMT

Hi.

Yes I'm using the P4 kernel. I'm also developing in C++.

The data and workspace are in new'ed arrays - which are not always 16 bit aligned. When I replaced the new with an _aligned_malloc the problem disappeared.

Thanks for your help.

David.

Re: Numerical differences in DGELSS between identical calls

TimP — Thu, 19 Aug 2004 03:41:02 GMT

David,

Yes, I think the MKL documents mention the possibility of better performance with aligned arguments, but perhaps not the possibility of slightly different round-off where mis-alignment requires more serial operations. Glad to hear that you found a reasonable solution.

Tim