[MAGEEC] [BEEBS] PCA characterisation of benchmarks & evaluation of features

James Pallister James.Pallister at bristol.ac.uk
Sun Jul 27 20:59:42 BST 2014


Hi Greg,

Sounds good!
> I believe this requires running the feature extractor across all the
> current benchmarks. If anyone can or has already done this and can
> send me the data it shall be appreciated. Otherwise I shall have a
> look into it Monday!
Nope, no-ones done this yet, I think. You'll have to clone and build the
the mageec repository, then compile the benchmarks with the mageec-gcc
plugin:

To build the mageec repository (with a temporary fix allowing it to be
built)
    git clone http://github.com/jpallister/mageec
    cd mageec
    git checkout temporary_fix
    mkdir build
    cd build
    ../configure
    make
    sudo make install

Use the plugin:
    gcc -fplugin=mageec test.c

This will output some stuff (which is likely to change as Simon works on
the plugin):
    Current Function: fn
      ( ft1) Basic Block Count     : 6
      ( ft2) BB with 1 successor   : 4
      ( ft3) BB with 2 successor   : 1
      ( ft4) BB with > 2 successor : 0
      ( ft5) BB with 1 predecessor : 4

With BEEBS you should be able to give this to make in CFLAGS, e.g. make
CFLAGS="-fplugin=mageec"


Hope this helps,
James

On 27/07/14 17:42, Greg Cawthorne wrote:
> Hello team!
>
> I have been working on doing some Principle Component Analysis of the
> features present in the current feature set. I have created a program
> which is ready to be used to evaluate beebs!
>
> Now I need a comprehensive set of feature vectors and which program &
> function they came from (as we're doing it on a function by function
> basis). Once I have this data I can perform PCA and then see which
> features are the most representative of the programs and also look to
> see if different types of benchmark have distinctly different
> important features.
>
> I believe this requires running the feature extractor across all the
> current benchmarks. If anyone can or has already done this and can
> send me the data it shall be appreciated. Otherwise I shall have a
> look into it Monday!
>
> Regards, Greg.
>
>
> _______________________________________________
> mageec mailing list
> mageec at mageec.org
> http://mageec.org/cgi-bin/mailman/listinfo/mageec

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mageec.org/pipermail/mageec/attachments/20140727/e904d1d8/attachment.html>


More information about the mageec mailing list