Search All of the Math Forum:
Views expressed in these public forums are not endorsed by
NCTM or The Math Forum.


Loesche
Posts:
16
Registered:
12/7/04


FIXED: [was] Again: Matlab for Linux and SVD
Posted:
Jun 5, 1997 4:39 AM


Hi there,
yesterday I posted a summary about problems with SVD under Matlab for Linux (see: Loesche SUMMARY: [was] Again: Matlab for Linux and SVD). Now it seems we found a quite nice workaround for that problem. Up to my knowledge the problems are caused by using an FPU which does its calculations in extended precision on one hand and Matlab using double precision all the time on the other.
Carlos Fonseca <fonseca@acse.shef.ac.uk> pointed out a possible workaround by using the function __setfpucw(unsigned short fpu_cw) which is documented/declared in /usr/include/i386/fpu_control.h. Thanks to Carlos for that. I wrote two short Cmex files. One for putting the FPU in double precision mode and another one for putting it back in extended mode. The source fpu_double.c is attached at the end.
Indeed, as the following transcript shows, this works  at least for my problem.
NNMP19:[loesche] >matlab < M A T L A B (R) > (c) Copyright 198495 The MathWorks, Inc. All Rights Reserved Version 4.2c Pedigree brian linus /rel/bld/matlab4/v4.2c/all/sandbox4/src Wed Apr 24 10:50:58 EDT 1996 NO FLEX Apr 24 1996 Commands to get started: intro, demo, help help Commands for more information: help, whatsnew, info, subscribe >> load C_22 >> format long >> svd(C_22) ??? Error using ==> svd Solution will not converge. >> norm(C_22) ??? Error using ==> norm Solution will not converge. >> fpu_double >> svd(C_22) ans = 1.00000000000000 1.00000000000000 1.00000000000000 1.00000000000000 1.00000000000000 0.99999999999999 >> norm(C_22) ans = 1.00000000000000 >> fpu_extended >> svd(C_22) ??? Error using ==> svd Solution will not converge. >> norm(C_22) ??? Error using ==> norm Solution will not converge. >> exit
As you see, when the FPU is in double precision mode, SVD and NORM converge. Switching back to extended precision shows the old behavior, i.e. SVD and NORM breaking down. When Matlab is terminated, the FPU precision mode is put to extended precision, regardless of what was used by Matlab.
Here fpu_double.c comes:
/* interface to routine __setfpucw in libc.so Ralf Loesche, 5. Juni 1997 */ #include <stdio.h> #include <math.h> #include <mex.h> #include <i386/fpu_control.h> /* prototype for external routine */ extern void __setfpucw(unsigned short fpu_cw); void mexFunction(int nargout,Matrix *argout[],int nargin,Matrix *argin[]) { unsigned short fpu_cw; fpu_cw = 0x127f; __setfpucw(fpu_cw); }
You obtain fpu_extended.c by changing fpu_cw = 0x127f; to fpu_cw = 0x137f; .
TMW could even put this source code on their solution search Web page for Solution Number: 5231. I think one can use fpu_double in the Matlab startup script without problems for the calculations done. Matlab uses double precision and does not take advantage of an fpu providing 80 bit digits and slightly higher precision.
Regards, Ralf Loesche
ps: I would like to see, how nonIntel FPU's act on this (like Cyrix 6x86, AMD K5/K6). With 100% Intel compatibility in mind, they should work the same. I don't have access to a nonIntel PC. Maybe someone can try that example.
pps: Even fpu_single can be done that way. However, I don't recommend using it ;)



