topic Re: Approximate Math Library in Intel® oneAPI Math Kernel Library

Approximate Math Library

capens__nicolas — Wed, 17 Jan 2007 14:14:50 GMT

Hi all,

I'm looking for the Approximate Math Library. It used to be at www.intel.com/design/Pentium4/devtools/but now I can't find it anymore. Does anyone know where it is now?

Thanks a lot,

c0d1f1ed

Re: Approximate Math Library

capens__nicolas — Sun, 21 Jan 2007 17:42:18 GMT

Found it: http://www.devmaster.net/forums/showthread.php?p=43580#post43580. Ideas for improvement are very welcome.

Re: Approximate Math Library

Intel_Software_Netw1 — Wed, 11 Apr 2007 20:57:06 GMT

Try this linkfor the 2.0 version. (Note, however, that this library is quite old -- the readmesays it's from 2000.)

Lexi S.

IntelSoftware NetworkSupport

http://www.intel.com/software

Re: Approximate Math Library

capens__nicolas — Thu, 12 Apr 2007 12:58:09 GMT

That's the same link as was given on the DevMaster.net forum. Indeed it's quite old, and I'm not very impressed by the performance / precision balance, I remembered it differentely... Anyway, I've written exp and log implementations myself which perform quite well. If you know more recent (free) open-sourcelibraries with SSE transcendental functions that would be great though. Thanks.

Re: Approximate Math Library

Intel_C_Intel — Wed, 23 May 2007 23:58:33 GMT

A very clever colleague of mine wrote this code before such functionality was added to the libraries from Intel. VML within MKL were intended to replace this library. VML has two variants - high and low accuracy with the high accuracy version have error bounds of 1 ulp (unit in the last place)while the low accuracy version has error bounds of 4 ulps. An even lower accuracy version,resulting in higher performance,is available in IPP. The functions ending in A11 have 11 bits of accuracy (approximately 12 bits of error). These may serve your purpose well.

Bruce