Re: [eigen] Bugs in Cholesky modules

[Andre Krause <post@xxxxxxxxxxxxxxxx>]

Gael Guennebaud wrote:
> Hi,
> 
> bugs fixed in rev 856821.
> 
> On Wed, Sep 3, 2008 at 10:20 PM, Timothy Hunter <tjhunter@xxxxxxxxxxxx> wrote:
>> Hello, I have found 2 bugs in the choleskyWithoutSquareRoot class:
>>
>> - it does not work with 1x1 matrices. I added a quick workaround on my local
>> copy that checks if the size is 1, but I do not know if there is a more
>> elegant solution than that.
> 
> I also did that, there is no better solution.
> 
> 
>> Also, how big can fixed-size matrices be? I encountered some problems to
>> compile code with Matrix<float, 1000, 1000> classes. (I am working on a more
>> complete bug report for that).
> 
> the limit is the maximal memory you can allocate on the stack, for instance:
> 
> float data[1000*1000]
> 
> won't work, so there is no way that Matrix<float, 1000, 1000>  works.
> Actually, from my experience there is absolutely no advantage to use
> fixed size matrices larger than 16.
> 


there are some important davantages!

1. allocation speed: stack allocation is incredibly faster than heap allocation
with new.

2. no memory fragmentation. due to the nature of the stack, the memory cannot
fragment...

you can try to increase the stack reserved for a program. this can be done by
custom compiler flags and settings.

for visual studio:

http://forums.microsoft.com/MSDN/ShowPost.aspx?PostID=979006&SiteID=1
"I would not use /F <num> as it only works if you compile and link in one step
(which is rare) - instead I would look at the linker /stack option.
Properties - Configuration Properties - Linker - System - Stack Reserve Size"



> Gael.
> 
>> Any help will be much appreciated.
>>
>> --
>>
>> Timothy Hunter
>>
>> Student (Stanford University)
>>
>> T. 404 421 3075
> 
> 
>